Paper deep dive
Generative AI-Based Data Augmentation for Oral Lesion Classification: The PhotoMOCI Dataset and Benchmark
Marco Parola, Mario G. C. A. Cimino, Sabrina Senatore
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Early detection of oral cancer via photographic imaging presents a promising avenue for large-scale oral cavity screening. However, the development of robust deep learning models is frequently hampered by the scarcity of high-quality, annotated datasets. To address this limitation, a novel and well-curated resource, the Photographic Multi-purpose Oral Cancer Imaging (PhotoMOCI) dataset, is introduced for developing models across multiple diagnostic tasks in oral oncology. Then, a comprehensive benchmark study was conducted to investigate how various data augmentation strategies influence the performance of image classifiers. Our analysis spans different generative AI frameworks, evaluating the efficacy of traditional methods against advanced generative approaches, including Generative Adversarial Networks (GANs) and Diffusion Models (DMs). Additionally, we propose the Synthetic Image Filter (SIF), a mechanism to select specific samples based on two auxiliary models: Synthetic Proxy Classifier to ensure samples are representative of the target class and Synthetic Image Detector to verify they appear realistic, thereby selecting only the high-utility images that contribute to improving downstream performance. Across the evaluated datasets and classifiers, the best SIF-filtered setup improves accuracy over traditional augmentation in all cases, with gains of +1.73% and +2.35% on PhotoMOCI and +2.38% and +2.08% on KOCD for ResNet50 and ViT, respectively. Our findings reveal that while the direct application of generative data augmentation may yield performance drops, the integration of SIF, considering (i) how synthetic data looks real and (ii) how it reflects the discriminative features of the belonging class, provides a simple yet effective mechanism to filter out synthetic samples that confuse the classifier during training.
Tags
Links
- Source: https://arxiv.org/abs/2609.35226v1
- Canonical: https://arxiv.org/abs/2609.35226v1
Trouble viewing inline? Open PDF directly →
Full Text
99,516 characters extracted from source content.
Expand or collapse full text
Generative AI-Based Data Augmentation for Oral Lesion Classification: The PhotoMOCI Dataset and Benchmark Marco Parola Email: mpar@create.aau.dk Affiliation: Aalborg University, Denmark Mario G.C.A. Cimino Email: mario.cimino@unipi.it Affiliation: University of Pisa, Italy Sabrina Senatore Email: ssenatore@unisa.it Affiliation: University of Salerno, Italy Abstract Early detection of oral cancer via photographic imaging presents a promising avenue for large-scale oral cavity screening. However, the development of robust deep learning models is frequently hampered by the scarcity of high-quality, annotated datasets. To address this limitation, a novel and well-curated resource, the Photographic Multi-purpose Oral Cancer Imaging (PhotoMOCI) dataset, is introduced for developing models across multiple diagnostic tasks in oral oncology. Then, a comprehensive benchmark study was conducted to investigate how various data augmentation strategies influence the performance of image classifiers. Our analysis spans different generative AI frameworks, evaluating the efficacy of traditional methods against advanced generative approaches, including Generative Adversarial Networks (GANs) and Diffusion Models (DMs). Additionally, we propose the Synthetic Image Filter (SIF), a mechanism to select specific samples based on two auxiliary models: Synthetic Proxy Classifier to ensure samples are representative of the target class and Synthetic Image Detector to verify they appear realistic, thereby selecting only the high-utility images that contribute to improving downstream performance. Across the evaluated datasets and classifiers, the best SIF-filtered setup improves accuracy over traditional augmentation in all cases, with gains of +1.73% and +2.35% on PhotoMOCI and +2.38% and +2.08% on KOCD for ResNet50 and ViT, respectively. Our findings reveal that while the direct application of generative data augmentation may yield performance drops, the integration of SIF, considering (i) how synthetic data looks real and (i) how it reflects the discriminative features of the belonging class, provides a simple yet effective mechanism to filter out synthetic samples that confuse the classifier during training. The PhotoMOCI dataset is publicly available on Kaggle at Parola (2025). keywordsData Augmentation, Oral Cancer, PhotoMOCI dataset, Synthetic Image Filter, Image Generation 1 Introduction Oral cancer (OC) remains a common and aggressive malignancy of the oral cavity, posing a substantial global burden of morbidity and mortality Kim and Kim (2020). This underscores the need for improved early recognition strategies Kim and Kim (2020); Di Fede et al. (2024), since, despite advances in treatment such as surgical resection and chemotherapy, the five-year survival rate remains around 60% Zhou et al. (2022). Consequently, photographic screening may support the early recognition and referral of suspicious oral lesions, while definitive diagnosis remains based on clinical assessment and, when indicated, histopathological examination Bramati et al. (2021); Ribeiro et al. (2022). An effective screening must balance accessibility, speed, and cost-efficiencyRibeiro et al. (2022). In this regard, recent research has explored the integration of Deep Learning (DL) with photographic imaging as a decision-support tool for oral lesion recognition and screening, rather than as a replacement for clinical or histopathological diagnosis Cimino et al. (2025); Parola et al. (2024). Since DL has shown considerable results in many medical image analyses, coupled with photographic imaging, it offers the potential to support clinical decision-making by providing diagnostic insights while reducing the cost related to other more expensive medical imaging techniques, such as X-ray, CT-scan, etc. Rajpurkar et al. (2022); Lee et al. (2021). The resulting performance therefore should not be interpreted as evidence of autonomous oral cancer diagnosis or as a substitute for the corresponding clinical and histopathological reference assessment. However, the development of robust DL-based systems is significantly constrained by the lack of large-scale, high-quality, and well-annotated datasets that are essential for training DL architectures Alzubaidi et al. (2023); Whang et al. (2022). Indeed, as highlighted in Section 3, oral cancer datasets for medical imaging are often limited in size due to challenges such as patient privacy concerns, the high cost and complexity of expert annotation, and the relatively low prevalence of early-stage lesions in clinical populations Alzubaidi et al. (2023); Whang et al. (2022); Di Fede et al. (2024). In this context, synthetic data generation emerges as a viable solution. By producing realistic images that increase visual variability within the photographic lesion domain, synthetic data can potentially augment existing datasets and improve the performance of DL models trained on them Tang and others (2023); Alzubaidi et al. (2023). The effectiveness of synthetic image generation as a data augmentation technique varies across different medical domains. While it has shown clear benefits in small-data scenarios, where real data are scarce, and model generalization is limited, it is not always consistently effective in broader or more data-rich applications Mantegna et al. (2025). In contrast, traditional transformation-based augmentation techniques often prove more effective and robust in large-data settings Mantegna et al. (2025), offering simpler and more computationally efficient improvements without the complexities of generative modeling. As discussed in the literature review (Section 2.1), synthetic data augmentation has been widely adopted in medical imaging and oral cancer. However, to the best of our knowledge, a systematic and comprehensive evaluation of synthetic data for photographic oral cancer imaging remains lacking. Existing studies typically focus on a single generative model, most often GAN-based approaches, or rely on a single dataset, which is frequently not publicly available, as highlighted in the dataset review in Section 3. Moreover, many contributions address alternative imaging modalities, such as histopathological images, limiting their applicability to photographic screening. To address both the limited availability of datasets and the absence of a comprehensive benchmark, we introduce a novel dataset designed to support future benchmarks and enable a more systematic evaluation of synthetic data generation for photographic oral cancer imaging. Specifically, the contributions of this work are: • Introducing a novel dataset in this scenario of data scarcity, the Photographic Multi-purpose Oral Cancer Imaging (PhotoMOCI) dataset, a well-curated collection for oral cancer that enables the development of models for multiple tasks. • Investigating the impact of different data augmentation strategies on image classification performance, and presenting a comprehensive benchmark across multiple datasets, image classifiers, generative models, and augmentation techniques. • Proposing the Synthetic Image Filter (SIF), a simple but effective strategy relying on two auxiliary classification problems to select specific synthetic images that contribute to improving the final classifier performance, based on how a synthetic image looks real and how it matches discriminative features of the belonging class. 2 Related work DL has achieved strong performance in medical imaging, but adoption is often limited by small datasets caused by privacy constraints, high annotation costs, and the rarity of certain diseases, leading to class imbalance and poor case representation Goodfellow and others (2014); Zhang and Qie (2023); Johnson and Khoshgoftaar (2019). Thus, in Section 2.1, we review traditional and synthetic data augmentation techniques that expand dataset size and diversity, including Generative Adversarial Networks and Diffusion Models, with a specific focus on conditioning mechanisms in Section 2.2. 2.1 Data augmentation technique review Traditional data augmentation. Traditional data augmentation increases training variability through predefined image transformations Krizhevsky et al. (2012), commonly categorized as geometric or photometric Tang and others (2023); Shorten and Khoshgoftaar (2019). Geometric transformations modify spatial structure (e.g., rotation, flipping, scaling), whereas photometric transformations alter pixel intensities (e.g., brightness, contrast) or add noise. These techniques, widely used in computer vision, have also proven effective in medical imaging applications Shin et al. (2016); Ronneberger et al. (2015); Wael Alosaimi (2022); Oya et al. (2023). Due to their computational efficiency Whang et al. (2022), traditional augmentations are broadly adopted and generally improve performance compared to training without augmentation Tang and others (2023); Krizhevsky et al. (2012); Ronneberger et al. (2015). However, they cannot generate truly novel samples or pathological features, as they only modify existing images Shorten and Khoshgoftaar (2019). This limits their effectiveness in scenarios with severe data scarcity, low diversity Alzubaidi et al. (2023), or class imbalance Johnson and Khoshgoftaar (2019); Kazemzadeh and others (2024), and they lack conditional generation capabilities. In contrast, DL-based generative approaches show greater promise in small-data medical imaging settings Mantegna et al. (2025), while traditional augmentation remains reliable for improving generalization in larger datasets Mantegna et al. (2025). Generative Adversarial Networks. Generative Adversarial Networks (GANs) Goodfellow and others (2014) represent the next generation of generative models after autoencoders (AEs) Hinton and Salakhutdinov (2006) and later Variational Autoencoders (VAEs)Kingma and Welling (2013). They introduced an adversarial training paradigm involving a generator and a discriminator. Through this process, the generator progressively learns to synthesize visually realistic images that deceive the discriminator Goodfellow and others (2014). Despite their success, GANs can suffer from training instability and mode collapse, where only a limited subset of the data distribution is generated Arjovsky et al. (2017); Tang and others (2023). Several variants addressed these issues, including Deep Convolutional GANs Radford et al. (2015), Wasserstein GANs Arjovsky et al. (2017), and StyleGAN Karras et al. (2019), which enables high-quality image synthesis and controllable attribute manipulation Karras and others (2020). GANs have been extensively applied in medical imaging for several generative tasks, including augmenting pathology contrast in mammography Rofena et al. (2025a); Rofena et al. (2025b), synthetic augmentation to improve: liver classification in CTFrid-Adar et al. (2018), covid-19 in chest X-rays Waheed et al. (2020); Neff et al. (2017), dermatology Qasim et al. (2021), and brain MRI classification Qi et al. (2020). In oral oncology, GANs have been used for oral cancer histology classification and segmentation Wael Alosaimi (2022); Dos Santos et al. (2023) and for fluorescence-based diagnosis Fujimoto et al. (2022). Conditional GANs further extend this framework by enabling class-specific data generation, particularly useful for rare conditions Mirza and Osindero (2014); Tang and others (2023). Diffusion Models. Diffusion Models (DMs) are the most recent class of generative models and have shown superior performance over previous approaches, especially in terms of image diversity Ho et al. (2020); Dhariwal and Nichol (2021); Tang and others (2023). DMs iteratively add Gaussian noise to training data and learn to reverse this process through denoising steps using UNet-based architectures Sohl-Dickstein et al. (2015); Ronneberger et al. (2015); Ho et al. (2020). During generation, samples are obtained by iteratively denoising random noise. A major drawback is the need for many denoising steps, resulting in longer inference times than single-pass methods Ho et al. (2020); Tang and others (2023). To reduce computational costs, Latent Diffusion Models (LDMs) perform diffusion in a compressed latent space learned via autoencoders Rombach et al. (2022). Additionally, Low-Rank Adaptation (LoRA) has been proposed as an efficient fine-tuning strategy, allowing diffusion models to adapt to new domains using minimal trainable parameters and reduced memory requirements Hu et al. (2021); Yang et al. (2024). Diffusion-based methods have been applied to address class imbalance in chest X-ray datasets Kazemzadeh and others (2024); Pinaya and others (2024) and dermatology images Sagers et al. (2023). In oral cancer, diffusion models have been explored for histopathological images Uliana and Krohling (2025), while their application to photographic images remains unexplored. 2.2 Condition mechanism in image synthesis Conditioning mechanisms guide generative models toward specific attributes and visual features Bourou et al. (2024). In oral cancer diagnosis, conditioning can support class-specific data generation. Conditioning in GANs. Conditioning extends GANs by incorporating auxiliary class label information c (e.g., one-hot class encoding, string class name) into the generator G and/or discriminator D. Given the GAN latent space z, basic conditioning approaches consist of concatenating c to z, which is effective for low-dimensional inputs such as class labels Bourou et al. (2024); Liu et al. (2020). Alternatively, Auxiliary Classifier GANs (AC-GANs) introduce an extra classification head in the discriminator, enforcing class label semantics using an additional supervised loss term Odena et al. (2017); Bourou et al. (2024). Conditioning in Diffusion Models. Diffusion Models support a wide range of conditioning signals, including text, labels, images, and segmentation masks Zhan et al. (2024). Conditioning can be integrated at different stages using operations such as concatenation, summation, or cross-attention. Text and class labels are typically embedded before being injected into the denoising process Zhan et al. (2024). Images or masks can directly guide the diffusion trajectory by constraining regions of interest or initializing the latent representation Ho et al. (2020). These mechanisms are particularly well suited to medical imaging, where precise control over generated content is critical Rombach et al. (2022); Kazemzadeh and others (2024). 3 Photographic Multi-purpose Oral Cancer Imaging Dataset Table 1: Existing photographic oral cancer datasets for computer vision. The column #Imgs (w/ lesions) reports total images and lesion‑positive ones (within round brackets). Name Year #Imgs (w/ lesions) Downstream tasks Kaggle OC Lips and TongueShivam and Prakrut (2020) 2020 131 (87) Binary image classification Roboflow OCsagari vijay (2021) 2021 323 (156) Multi-label image classification, Object detection, Semantic segmentation Mendeley Oral Images Islam et al. (2023) 2023 373 (158) Binary image classification KOCD Rashid (2024) 2024 950 (500) Binary image classification PhotoMOCI (ours) 2026 700 (700) Multi-label image classification, Object detection, Semantic segmentation, CBR In this section, we review current public data resources and their limitations. Then, we introduce a novel dataset: Photographic Multi-purpose Oral Cancer Imaging. We review publicly available oral cancer (OC) photographic datasets released between 2020 and 2025, excluding other imaging modalities, such as histopathology and radiography. Despite the growing number of works on OC imaging, the availability of datasets remains a major limitation. Many studies rely on private or poorly documented datasets, primarily due to patient privacy concerns and institutional constraints, which significantly hinder reproducibility and fair methodological comparison. Indeed, several datasets are accessible only upon request and require detailed justification from potential users Rabinovici-Cohen et al. (2024); Lee et al. (2023); Lin et al. (2021); Jubair et al. (2022); Welikala et al. (2021); Heo et al. (2022); Warin et al. (2022), resulting in a time-consuming and often discouraging process. Even more restrictive are works in which datasets are not accessible at all, either because they were never intended for public release or due to legal and ethical restrictions Kouketsu et al. (2024); Vinayahalingam et al. (2024). Table 1 provides an overview of publicly available photographic datasets sorted by year, detailing size and the supported computer vision downstream tasks. Such analysis highlights several limitations in the datasets used for oral healthcare research. First, all publicly available oral cancer datasets contain fewer than 1,000 samples, making it difficult to generalize findings. Second, the number of samples further decreases when annotations are provided for multiple downstream tasks, such as classification, detection, or segmentation. Indeed, three out of five datasets are curated only for binary classification (cancer vs. non-cancer). The only dataset supporting multiple downstream tasks comprises just 323 images. Finally, note that many publicly available datasets appear larger than they are, as they include the ”healthy” class. These images are easier to collect because they do not require waiting for rare pathological events. Consequently, they are often used to artificially inflate dataset sizes, potentially overselling the actual clinical contribution while the volume of high-value pathological samples remains limited. This scarcity of well-annotated, publicly accessible datasets underscores the need for greater efforts to create and share resources that support robust and reproducible research in healthcare. With this premise, we introduce the Photographic Multi-purpose Oral Cancer Imaging dataset, a multipurpose dataset designed for cancer recognition in oral cavity photographic imaging. It supports a wide range of downstream tasks (some of which were previously addressed), including major computer vision problems such as image classification, object detectionParola et al. (2023), and semantic segmentationParola et al. (2025a), as well as clinically oriented applications like case-based reasoning Parola et al. (2024); Kenny and Keane (2019); Cimino et al. (2025); Keane and Kenny (2019). By offering a dataset combining technical and clinical needs, PhotoMOCI supports research on photographic oral cancer screening. Nevertheless, this work specifically focuses on image classification. The study adhered to the ethical standards outlined in the Declaration of Helsinki and its amendments for research involving human participants. Ethical approval was obtained from the local ethics committee (n.12 07/05/2024). Written informed consent was obtained from all participants following established guidelines for conducting scientific research. Between 2021 and 2025, photographic images of the oral cavity were collected from patients attending medical examinations at the Oral Medicine Unit of the P. Giaccone University Hospital in Palermo, Italy. Dental hygienists and oral medicine consultants acquired these images using a Nikon D7200 digital camera equipped with a 105 m lens and an SB-R200 macro flash, offering an inexpensive alternative to advanced imaging systems. The resulting image dataset is composed of high-quality resolution ranging from ∼ 2500x2000 to ∼ 6000x4000 pixels. Additionally, histopathological analysis was performed to confirm the diagnosis and subsequent labeling. The final version includes 700 images belonging to three specific pathologies: aphthous, traumatic, and neoplastic (cancer class). The majority of images in the PhotoMOCI dataset feature a single lesion; however, 22 images contain multiple instances of lesions within the same photo for a total of 730 oral lesions (253 aphthous, 257 traumatic, and 220 neoplastic). Note that there are no instances where lesions from different classes coexist within the same image. All cases involving multiple lesions pertain to a specific scenario: the multiple aphthous lesions. The dataset contains no more than one image from the same patient. (a) Sex distribution. (b) Age distribution. Figure 1: PhotoMOCI metadata distribution by sex in (a) and age group in (b), stratified by pathology class. Figure 1 summarizes the PhotoMOCI metadata distribution by sex and age group, with each bar split according to the three pathology classes. The sex distribution shows a small mismatch, with more men than women; however, across the diagnostic classes, the samples remain balanced. Regarding age, most individuals involved in the data collection are between 30 and 60 years old. In the younger segment (10–20 years), the neoplastic class is rare and traumatic lesions dominate, whereas neoplastic lesions occur more frequently among older individuals. Other commonly reported clinical metadata, such as anatomical site and pathological stage, and visual lesion features, such as lesion size, border color, and lesion color, are not available for PhotoMOCI. For the core computer vision tasks, the images were annotated using the open-source tool COCO Annotator Brooks (2019), generating a dataset in the standard COCO format for computer vision applications. Initial annotation of lesion regions was performed manually by a dental intern, who delineated lesion boundaries and assigned pathology labels returned by histopathological analysis. Subsequently, all annotated images were reviewed and refined by two senior clinicians with more clinical experience. The revisions focused exclusively on segmentation boundaries to improve their accuracy, as the labels are based on the histopathological results. The bounding box is then generated by considering the minimum rectangle that encloses the lesion segment. The images and the json annotation files can be accessed on Kaggle Parola (2025). 4 Methodology Figure 2: Overview of the proposed methodology for oral lesion classification. In the ochre dotted box, the Synthetic Image Generation based on generative models. In the green dotted box, the Synthetic Image Filter combining the Synthetic Proxy Classifier and the Synthetic Image Detector to select informative synthetic samples. In the light-blue dotted boxes, three training setups: using the original small dataset (top), the augmented synthetic dataset (middle), and the augmented synthetic filtered dataset (bottom), the latter aiming to improve generalization. The proposed methodology follows a three-stage pipeline designed to generate, select, and exploit synthetic data for oral cancer classification, as illustrated in Figure 2 and detailed in Sections 4.1, 4.2, and 4.3, respectively. • Synthetic Image Generation (ochre dotted box). This stage relies on a generative model fine-tuned on an OC dataset (usually small) to generate synthetic images. • Synthetic Image Filter (green dotted box). This stage exploits the proposed SIF to select specific synthetic images that, coupled with the original dataset, improve the oral lesion classifier. • Training the final oral cancer classifier (ligh-blue dotted box). This stage is responsible for training the oral lesion classifier under different setups. 4.1 Synthetic Image Generation Let real=(xi,yi)i=1ND_real=\(x_i,y_i)\_i=1^N be the Small Real Dataset, where xi∈x_i denotes the input real image belonging to image space X and yi∈y_i its corresponding ground‑truth label from a predefined set Y. Let synth=(xj,yj)j=1MD_synth=\(x_j,y_j)\_j=1^M be the Synthetic Dataset, where xj∈x_j is a synthetic image and yj∈y_j the corresponding label derived from the generation process. The Synthetic Image Generation stage is responsible for generating synthD_synth, relying on a Synthetic Image Generator denoted as g(⋅)g(·) finetuned on realD_real. For the generation of oral lesions, all synthetic samples (xj,yj)∈synth(x_j,y_j) _synth are generated from g(.)g(.) using conditioning information c (whether a string, a vector, or a single value) referring to class labels yiy_i. Therefore, both xjx_j and yjy_j result as a function of c. Specifically, we employ two state-of-the-art generative models. The first is Stable Diffusion Rombach et al. (2022), a latent diffusion framework fine-tuned through the LoRA domain adaptation technique. The second is StyleGAN3 Karras et al. (2021), which introduces an alias-free generator replacing conventional upsampling and filtering operations with low-pass filters. For each generative model, we explore two conditioning strategies, leading to four distinct synthetic generation setups: • SD(txt): Stable Diffusion conditioned on text. Lesion descriptions are first encoded into embeddings using CLIP Radford et al. (2021). These embeddings are then combined with the latent representations during the denoising process by summing them, allowing the model to generate images that closely align with the provided textual descriptions. • SD(txt+img): Stable Diffusion conditioned on both text and images. Similar to the previous setup, textual embeddings are fused with image embeddings. However, instead of executing the full denoising trajectory from pure noise, the process is initialized from a conditioning image and proceeds for a limited number of denoising steps. This approach leverages the input image as a structural prior. Figure 3 illustrates this text–image conditioning scheme for a neoplastic lesion, where the conditioning image is injected at different stages of the denoising process. • SG3(lbl): StyleGAN3 conditioned on class labels. Here, labels are one-hot encoded and projected into a learned embedding space. The resulting embeddings are then concatenated with the generator’s latent codes, enabling label-aware synthesis. • AC-SG3: Auxiliary Classifier StyleGAN3 conditioned on class labels. In this configuration, conditioning is enforced through the discriminator, which is augmented with an auxiliary classifier head that predicts the class label in addition to the adversarial objective, thereby encouraging the generator to generate real images that are not only indistinguishable from real images but also belong to the specified class. 0 1 2 3 4 5 6 7 8 9 10 20 30 40 50 60 70 80 90 100 10 20 30 40 50 60 70 80 90 100 Step Step (a) (b) Figure 3: Two examples of diffusion processes using SD(txt+img) with conditioning mechanisms based on text and image for neoplastic and traumatic lesions in (a) and (b) from PhotoMOCI, respectively. In each example, ten denoising processes with image conditioning performed at different steps. The red-bordered images indicate where the conditioning image is introduced. In the last row, with water-green borders, the original conditioning image injected at the 100th step, overriding earlier ones. Figure 3 presents two examples of diffusion processes in the SD(txt+img) setup. The model is a Stable Diffusion architecture finetuned on the PhotoMOCI dataset using LoRA over 100 denoising steps. In each case, the same conditioning image, belonging to the neoplastic Figure 3(a) and traumatic Figure 3(b) classes, is injected at different stages of the denoising process to guide generation. The specific step at which the conditioning image is introduced is highlighted with a red border. The final image in the last row, marked with a water-green border, displays the original conditioning image injected at the final denoising step, which overrides the preceding steps and fully dominates the output. 4.2 Synthetic Image Filter While generative models can synthesize large volumes of data, the utility of such samples is not always significant. Poorly generated samples, characterized by artifacts or semantic misalignment with the target class, introduce label noise that can degrade the performance of downstream classifiers. To mitigate this, we propose the Synthetic Image Filter (SIF), a mechanism designed to select only those samples (xj,yj)(x_j,y_j) from synthD_synth that enhance the generalization of the final oral cancer classifier. As illustrated in Figure 2 (green dotted box), SIF implements a filtering mechanism based on two auxiliary models trained for complementary purposes: a Synthetic Proxy Classifier (SPC), denoted as fSPCf_SPC, and a Synthetic Image Detector (SID), denoted as fSIDf_SID. The fSPCf_SPC is a classification model trained exclusively on synthetic images to solve the same downstream classification task as the oral lesion classifier detailed in the next section. Intuitively, this model evaluates whether a synthetic image xjx_j exhibits visual patterns coherent with its intended label yjy_j. On the other hand, SID (fSIDf_SID) is a binary classifier trained to discriminate real images xi∈realx_i _real from synthetic images xj∈synthx_j _synth. Its objective is to capture the distributional shift between synthetic and real images, providing a machine-based measure of realism for new images x x: fSID(x^)∈[0,1]f_SID( x)∈[0,1] (1) where a value closer to 11 indicates high confidence that the sample is synthetic. In this context, “realism” refers to machine indistinguishability from the real-image distribution as estimated by SID, and should not be interpreted as clinical realism or diagnostic plausibility assessed by clinicians. Although these notions may be related, since clinically implausible artifacts can make synthetic samples easier to detect, SID evaluates realism only from the perspective of the trained discriminator. Finally, SIF combines the predictions of these two auxiliary models to select synthetic images for data augmentation. Specifically, a sample is retained only if it is correctly classified by SPC, indicating class label semantic consistency, and misclassified by SID, suggesting strong visual similarity to real images. The resulting Filtered Dataset filtD_filt is formally defined as filt=(xj,yj)∈synth∣fSPC(xj)=yj∧fSID(xj)<τD_filt=\(x_j,y_j) _synth f_SPC(x_j)=y_j f_SID(x_j)<τ\, where τ is a decision threshold (we adopt τ=0.50.5). The selected samples are subsequently combined with the original small real dataset to form the final training set real∪filtD_real _filt. 4.3 Training the oral lesion classifier In the last stage, the oral lesion classifier is trained under different data configurations to assess the impact of synthetic data and the proposed filtering strategy. Specifically, the classifier is trained using one of the following input datasets: • the original small real dataset realD_real, which may generate a classifier that generalizes poorly because it is trained on a limited amount of data; • the original dataset augmented with the full Synthetic Dataset real∪synthD_real _synth, which may produce a classifier with performance drops due to the uncontrolled integration of noisy synthetic images; • the original dataset augmented with the Filtered Dataset real∪filtD_real _filt, which produce a more robust classifier. 5 Experiments and results The experiment section is organized into two parts: The first part, presented in Section 5.1, describes how we finetune different generative architectures and reports experiments on synthetic image generation. Then, we describe how we develop the auxiliary classifiers composing SIF and how it is used to train oral lesion classifiers under different synthetic data augmentation setups in Sections 5.2 and 5.3, respectively. Finally, we conduct an ablation study to understand the impact of the different SIF components on the overall filter. The experiments are implemented in PyTorch 1.10.2 (CUDA 11.3) and are publicly available on GitHub at Parola et al. (2025b). They are run on NVIDIA A100 GPUs with 32GB of memory. Datasets. We evaluate our approach on PhotoMOCI for the multi-class problem, and on KOCD (version 1, released under Apache 2.0) Rashid (2024) for the binary problem, since they represent the largest and most recent datasets available in this field. For experimental consistency and reproducibility, all experiments on PhotoMOCI follow the official dataset partition, consisting of 490, 105, and 105 images for the training, validation, and test sets, available at Parola (2025). The KOCD dataset consists of 950 images labeled according to a binary classification setup: 450 cancer and 500 non-cancer. The images are collected from 950 patients, with each patient contributing one image; therefore, no near-duplicate images are present in the dataset. Since KOCD is released without predefined partitions, we define a hold-out 70%-15%-15% split, resulting in 665 (321 cancer and 344 non-cancer), 142 (64 cancer and 78 non-cancer), and 143 (66 cancer and 77 non-cancer) images for the training, validation, and test sets, respectively. The split is generated with a fixed random seed and is therefore reproducible across all experiments Parola et al. (2025b). For both PhotoMOCI and KOCD, the same splits are consistently used throughout all the experiments. 5.1 Synthetic Image Generation Configuration Setup. We evaluated four distinct synthetic image generation setups on 100 samples per class for each epoch against the real data. Alongside such generation setups, we established a traditional augmentation baseline consisting of both geometric and photometric transformations: random flipping, color jittering, and affine transformations (rotations ±10∘± 10 , translations ±15%± 15\%, scaling 85-115%, and shearing ±5∘± 5 ). The quality of all synthetic sets was benchmarked against the original, non-augmented data. Generative models were finetuned on the training set and conditioned with dataset-specific prompts: ”a medical photograph of an ’x’ lesion” for PhotoMOCI, where ’x’ is the class label (aphthous, traumatic, neoplastic), and ”a medical photograph of a ’x’ tissue” for KOCD, where ’x’ was either cancerous or healthy (cancer and non-cancer classes). Image-conditional setups in the case of SD(txt+img), used images from the training set. Since image conditioning can occur at different stages of the diffusion process, as shown in Figure 3, we empirically identified a range of steps that balance fidelity to the conditioning image with sufficient diversity in the generated output. Specifically, it is applied at a random step between steps 55 and 75 of the 100-step diffusion process. For the training of our conditional and AC-StyleGAN3 variants, we adopted the ‘stylegan3-t‘ configuration and trained the models from scratch. With a batch size of 16 images, the training process stabilized with R1 regularization and a γ coefficient of 0.6. Distinct learning rates were set for the generator (2.5×10−32.5× 10^-3) and the discriminator (2.0×10−32.0× 10^-3). To prevent discriminator overfitting, we employed Adaptive Discriminator Augmentation (ADA) alongside real-time horizontal mirroring. For the training of our latent diffusion models, we fine-tuned the Stable Diffusion v1.5 model using Low-Rank Adaptation (LoRA). The model was trained for a maximum of 200 epochs with a batch size of 32. We monitored the validation loss, ‘val/loss‘, to save the top three checkpoints. An 8-bit Adam optimizer was employed with a learning rate of 1.0×10−41.0× 10^-4 for the U-Net and 0.5×10−40.5× 10^-4 for the text encoder, using a weight decay of 1.0×10−21.0× 10^-2. The LoRA configuration featured a rank of 16, an alpha of 32, and a dropout of 0.15, with adapters applied to the attention and projection modules of the U-Net and text encoder, respectively. Training was optimized using ‘bf16-mixed‘ precision, 4-bit quantization. Metrics. To evaluate the performance of generative models, we employ different established state-of-the-art metrics Jalali et al. (2023). Such evaluation metrics can be categorized into two main groups. The first comprises distance-based metrics, including the Fréchet Inception Distance (FID) Heusel et al. (2017) and Kernel Inception Distance (KID) Bińkowski et al. (2018) scores, which quantify the similarity between the distributions of generated and real data. The second group includes quality-based metrics that assess the fidelity and variability of the class labels predicted by a pre-trained classifier. This includes Precision and Recall Kynkänniemi et al. (2019), and Coverage Naeem et al. (2020) scores. Table 2: Synthetic Image Generation results on PhotoMOCI. Synthetic img. gen. setup FID↓ KID↓ Prec.↑ Recall↑ Cov.↑ Tr. aug. 230.30 0.126 0.317 0.183 0.058 SD(txt) 223.92 0.079 0.149 0.660 0.191 SD(txt+img) 179.24 0.072 0.543 0.319 0.521 SG3(lbl) 195.77 0.099 0.404 0.234 0.287 AC-SG3 290.42 0.230 0.426 0.011 0.032 Table 3: Synthetic Image Generation results on KOCD. Synthetic img. gen. setup FID↓ KID↓ Prec.↑ Recall↑ Cov.↑ Tr. aug. 262.91 0.126 0.284 0.632 0.242 SD(txt) 225.38 0.074 0.273 0.762 0.231 SD(txt+img) 143.87 0.034 0.867 0.573 0.888 SG3(lbl) 169.57 0.062 0.888 0.112 0.748 AC-SG3 319.78 0.245 0.035 0.035 0.042 Results. Tables 2 and 3 report the synthetic image generation task results on PhotoMOCI and KOCD, respectively. Across both datasets, consistent trends emerge: SD(txt-img) generally outperforms the other approaches in most metrics, except Recall on PhotoMOCI and both Precision and Recall on KOCD. On PhotoMOCI, SD(txt) achieves higher Recall, indicating that, in the context of diffusion models, text-only conditioning leads to greater diversity in the generated samples, effectively acting as a “general-purpose” generator. However, this diversity negatively affects the quality, as indicated by its low precision value, meaning that many of the generated images are not realistic. On KOCD, SD(txt) again surpasses SD(txt-img) in Recall, while SG3(lbl) achieves the best Precision, with its capacity to generate images that, from the perspective of a pre-trained classifier, closely reflect the label distributions observed in real samples. StableDiffusion (txt) StableDiffusion (txt+img) traumatic aphthous neoplastic StyleGAN3 (lbl) AuxiliaryClassifier-StyleGAN3 traumatic aphthous neoplastic Figure 4: Synthetic image examples generated from the PhotoMOCI dataset using four synthetic image generation setups: Stable Diffusion conditioned on text (top-left), Stable Diffusion conditioned on text and image (top-right), StyleGAN3 conditioned on text (bottom-left), and AC-SG3 (bottom-right). Within each block, the first, second, and third rows correspond to traumatic, aphthous, and neoplastic lesions, respectively. StableDiffusion (txt) StableDiffusion (txt+img) cancer non-cancer StyleGAN3 (lbl) AuxiliaryClassifier-StyleGAN3 cencer non-cancer Figure 5: Synthetic image examples generated from the KOCD using four synthetic image generation setups: Stable Diffusion conditioned on text (top-left), Stable Diffusion conditioned on text and image (top-right), StyleGAN3 conditioned on text (bottom-left), and AC-SG3 (bottom-right). Within each block, the first row corresponds to the cancer class, and the second row corresponds to the non-cancer class. This insight is also supported by Figures 4 and 5, which present examples of generated synthetic images by class with the considered generation strategies on PhotoMOCI and KOCD, respectively. Qualitatively, the images produced by SD(txt) are clearly recognizable as synthetic to the human eye, showing anatomical features of the oral cavity that do not exist in either dataset and exhibiting almost no variability within the same class. By contrast, SD(txt-img) and SG3(lbl) generate synthetic images that appear visually plausible and capture the distinctive features of the classes (e.g., necrotic tissue in the case of neoplastic lesions) in both datasets. Finally, a discrepancy can be observed in the quality of the images produced by AC-SG3 across datasets: on PhotoMOCI, the generated samples lack intra-class variability, displaying a recurring visual pattern within the same class, whereas this phenomenon is not present on KOCD. This suggests that AC-SG3 may struggle to generalize across settings, such as PhotoMOCI, where the class granularity is higher. Finally, we generate the synthetic samples that will be used in the next stages of the methodology. Specifically, for each synthetic image generation setup trained on KOCD, we generate 3000 images (1500 per class). While we generate 2100 images for PhotoMOCI (700 per class). We report the wall-clock time required by each generative setup in Table 4, referring to the same hardware, an NVIDIA A100 GPU with 32GB of memory. For each setup, the generation of a single image was repeated 10 times and we report the mean generation time with the 95% confidence interval. Additionally, for each dataset, we report the actual time for the entire synthetic image generation process. GAN-based models are computationally more efficient than diffusion-based models; however, synthetic image generation is performed only once as an offline preprocessing step and therefore does not affect the downstream classifier training and inference phases. Table 4: Computational time for generating synthetic images on an NVIDIA A100 GPU. Synthetic img. gen. setup Single image time (s) PhotoMOCI 2100 images KOCD 3000 images SD(txt) 1.86 ± 0.05 1 h 05 min 06 s 1 h 33 min 00 s SD(txt+img) 0.82 ± 0.06 28 min 42 s 41 min 00 s SG3(lbl) 0.045 ± 0.003 1 min 35 s 2 min 15 s AC-SG3 0.048 ± 0.004 1 min 41 s 2 min 24 s Table 5: SIF modules performance: classification results of SPC and SID on PhotoMOCI and KOCD considering the four synthetic data generation. Dataset Synthetic img. gen. setup Accuracy SPC ↑ SID↓ PhotoMOCI (ours) SD(txt) 0.368 ± 0.045 0.788 ± 0.032 SD(txt+img) 0.771 ± 0.037 0.643 ± 0.040 SG3(lbl) 0.740 ± 0.029 0.596 ± 0.038 AC-SG3 0.893 ± 0.036 0.921 ± 0.036 KOCDRashid (2024) SD(txt) 0.617 ± 0.025 0.709 ± 0.030 SD(txt+img) 0.914 ± 0.029 0.677 ± 0.021 SG3(lbl) 0.857 ± 0.033 0.780 ± 0.024 AC-SG3 0.783 ± 0.026 0.893 ± 0.019 5.2 Synthetic Image Filter Configuration Setup. For the implementation of the SIF auxiliary classifiers, SPC and SID, we employ two ResNet50 architectures. For each setup, we perform a training hyperparameter grid search. With the batch size maximized to 64 under GPU constraints, we search lrlr in [1e–4,5e–4,1e–5,5e–5,1e–6,5e–6][1e 4,5e 4,1e 5,5e 5,1e 6,5e 6] using the Adam optimizer Kingma (2014). Given the best configuration, models were trained for 150 epochs with early stopping (patience 10). To get statistical results, we repeated each training 10 times with different seeds, reporting the mean and 95% confidence interval. Regarding SPC, we use 490 and 665 synthetic images for training on PhotoMOCI and KOCD, respectively; while validation and test sets include only the original real images. For SID, we use the original train, validation, and test splits and augment each with an equal number of randomly selected synthetic images, preserving class balance to train a binary classifier able to distinguish between real and synthetic. This results in 1400 and 1900 for PhotoMOCI and KOCD, respectively. Results. Table 5 reports the results obtained on both classification tasks by SPC and SID on PhotoMOCI and KOCD. We observe that, for SPC addressing the same oral lesion classification task, the accuracy is high across all configurations, except for SD(txt) on PhotoMOCI, suggesting that the class distributions are generally well separable. In contrast, for SID, we would ideally expect performance close to random guessing; however, the observed accuracy values are higher, indicating a discrepancy between the distributions of real and synthetic data. This discrepancy is not necessarily apparent to human inspection but is captured through latent patterns. Consistent with Tables 3 and 2, SD(txt+img) and SG3 achieve the highest performance, suggesting that the effectiveness of the two auxiliary classification tasks can be preliminarily assessed using generative evaluation metrics (FID, KID, Precision, Recall, and Coverage). Table 6: Number and percentage of synthetic images retained by SIF on PhotoMOCI and KOCD, reported in aggregate and by class. Synthetic img. gen. setup PhotoMOCI KOCD Total Traumatic Aphthous Neoplastic Total Cancer Non-cancer SD(txt) 139 (6.6%) 30 (4.3%) 83 (11.9%) 26 (3.7%) 308 (10.3%) 91 (6.1%) 217 (14.5%) SD(txt+img) 470 (22.4%) 157 (22.4%) 119 (17.0%) 194 (27.7%) 653 (21.8%) 362 (24.1%) 291 (19.4%) SG3(lbl) 367 (17.5%) 141 (20.1%) 137 (19.6%) 89 (12.7%) 776 (25.9%) 415 (27.7%) 361 (24.1%) AC-SG3 70 (3.3%) 12 (1.7%) 19 (2.7%) 39 (5.6%) 152 (5.1%) 24 (1.6%) 128 (8.5%) After training the SIF modules, we apply the complete filtering pipeline to the synthetic images generated by each configuration and retain only the samples accepted by both auxiliary classifiers. Table 6 reports the number and percentage of synthetic images retained by the SIF for each synthetic generation setup for PhotoMOCI and KOCD (both overall and per class). Percentages are computed with respect to the generated pool of each setup: 2100 images for PhotoMOCI (700 per class) and 3000 images for KOCD (1500 per class). It provides an overview of how many generated images effectively contribute to the augmented training sets. Overall, the number of retained samples varies considerably across setups, with SD(txt+img) and SG3(lbl) generally producing the largest number of accepted samples, with 470 (22.4%) and 367 (17.5%) on PhotoMOCI and 653 (21.8%) and 776 (25.9%) on KOCD, respectively. Moreover, the retained images are reasonably distributed across disease categories rather than being concentrated in a single class. Although some variability is observed, no pronounced bias toward or against a particular pathology emerges after filtering. The computational time required by SIF corresponds to the inference time of the auxiliary architectures used in the filtering pipeline. We average 10 inference runs for each classification architecture. ResNet50-based SIF filtering requires 10.1 ms per image, while ViT-based filtering requires 21.7 ms. This introduces a minimal overhead and does not impact the overall computational cost. Table 7: Classification performance on PhotoMOCI using ResNet and ViT under different data augmentation strategies. The evaluated setups include training without augmentation (No aug.) and with traditional augmentation (Tr. aug.), marked with ✗ since SIF is not applicable, as well as four generative-AI–based setups evaluated both without and with SIF, distinguished by ✓. Highlighted in gray, the setups where SIF improves performance over non-SIF counterparts and surpasses traditional augmentation. Highlighted in bold, the best-performing augmentation setup for each classifier. Model Conditioning SIF Acc Prec Recall F1-score ResNetHe et al. (2016) No aug. ✗ 0.81050± 0.00532 0.81538± 0.00553 0.81076± 0.00581 0.81414± 0.00601 Tr.aug.(base) ✗ 0.82519± 0.00600 0.83130± 0.00572 0.82316± 0.00603 0.82873± 0.00612 SD(txt) 0.74153± 0.00590 0.74797± 0.00619 0.74225± 0.00528 0.74550± 0.00493 ✓ 0.80076± 0.00544 0.82653± 0.00485 0.80567± 0.00521 0.81237± 0.00495 SD(txt+img) 0.83655± 0.00520 0.83704± 0.00446 0.83679± 0.00499 0.83698± 0.00524 ✓ 0.84253± 0.00498 0.85137± 0.00507 0.84294± 0.00481 0.85095± 0.00511 SG3(lbl) 0.83282± 0.00488 0.84001± 0.00519 0.83335± 0.00535 0.83560± 0.00515 ✓ 0.83603± 0.00454 0.84261± 0.00478 0.83776± 0.00440 0.83860± 0.00423 AC-SG3 0.79153± 0.00638 0.80797± 0.00533 0.79225± 0.00604 0.79450± 0.00592 ✓ 0.80988± 0.00545 0.81530± 0.00630 0.80849± 0.00577 0.81069± 0.00563 ViTDosovitskiy et al. (2021) No aug. ✗ 0.82173± 0.00481 0.82330± 0.00499 0.82033± 0.00508 0.82125± 0.00510 Tr.aug.(base) ✗ 0.82950± 0.00617 0.83176± 0.00628 0.82984± 0.00637 0.83014± 0.00615 SD(txt) 0.75800± 0.00534 0.76146± 0.00593 0.75724± 0.00520 0.76031± 0.00563 ✓ 0.82109± 0.00467 0.82739± 0.00446 0.82335± 0.00489 0.82616± 0.00423 SD(txt+img) 0.84500± 0.00497 0.85190± 0.00497 0.84892± 0.00489 0.84876± 0.00473 ✓ 0.85296± 0.00414 0.85948± 0.00427 0.85310± 0.00416 0.85761± 0.00444 SG3(lbl) 0.84174± 0.00511 0.84655± 0.00519 0.84281± 0.00532 0.84502± 0.00531 ✓ 0.84741± 0.00413 0.85269± 0.00432 0.84444± 0.00429 0.84897± 0.00435 AC-SG3 0.81453± 0.00580 0.82138± 0.00554 0.81724± 0.00535 0.81882± 0.00500 ✓ 0.82778± 0.00440 0.83003± 0.00448 0.82721± 0.00453 0.82959± 0.00416 5.3 Training Oral classifier Configuration Setup. The recognition of lesions from photographic datasets was addressed as a classification task. We studied the impact of synthetic data on performance by training models on both original and augmented datasets (PhotoMOCI and KOCD). For each dataset, five augmentation setups were compared against a baseline trained only on the original data: traditional augmentation, two diffusion-based augmentations, and two GAN-based augmentations. Except in the traditional augmentation setup, which was applied dynamically during training following common practice, synthetic images were added to the originals for training with and without applying SIF. As image classifiers, we considered both a convolutional model and a transformer: ResNet50 He et al. (2016) and a Vision Transformer (ViT) Dosovitskiy et al. (2021), both pre-trained on ImageNet Deng et al. (2009) from torchvisionThe model’s evaluation is based on standard classification metrics: accuracy, precision, recall, and F1-score. For each model and setup, we perform a hyperparameter grid search. With batch=64batch=64, we search the best lrlr in [1e–4,5e–4,1e–5,5e–5,1e–6,5e–6][1e 4,5e 4,1e 5,5e 5,1e 6,5e 6] using Adam Kingma (2014). As for the experiments in the previous section, we train the models for 150 epochs with an early stopping condition, repeating the training, changing the seed, and reporting the mean and 95% confidence interval. Results. Tables 7 and 8 summarize classification metrics for the PhotoMOCI and KOCD datasets, respectively, showing both similarities and differences. It is important to note that the datasets are relatively small compared to standard benchmarks such as ImageNetDeng et al. (2009) or COCOLin et al. (2014). In all experiments, the best-performing setup involves Stable Diffusion conditioned on both image and text, utilizing either ResNet or ViT when applying SIF. On the PhotoMOCI dataset, this setup improves accuracy by 1.73% and 2.35% over traditional augmentation (baseline) for ResNet and ViT, respectively. On KOCD, these improvements over the baseline reach 2.38% and 2.08%. Table 8: Classification performance on KOCD using ResNet and ViT under different data augmentation strategies. The evaluated setups include training without augmentation (No aug.) and with traditional augmentation (Tr. aug.), marked with ✗ since SIF is not applicable, as well as four generative-AI–based setups evaluated both without and with SIF, distinguished by ✓. Highlighted in gray, the setups where SIF improves performance over non-SIF counterparts and surpasses traditional augmentation. Highlighted in bold, the best-performing augmentation setup for each classifier. Model Conditioning SIF Acc Prec Recall F1-score ResNetHe et al. (2016) No aug. ✗ 0.90935± 0.00453 0.90028± 0.00452 0.90074± 0.00449 0.90026± 0.00454 Tr.aug.(base) ✗ 0.93109± 0.00554 0.93103± 0.00547 0.93120± 0.00563 0.93095± 0.00576 SD(txt) 0.88564± 0.00637 0.89160± 0.00656 0.88713± 0.00625 0.89121± 0.00700 ✓ 0.91616± 0.00410 0.91937± 0.00474 0.91668± 0.00422 0.91871± 0.00390 SD(txt+img) 0.95011± 0.00498 0.95225± 0.00501 0.94917± 0.00533 0.95075± 0.00511 ✓ 0.95492± 0.00467 0.95719± 0.00501 0.95266± 0.00461 0.95543± 0.00521 SG3(lbl) 0.94013± 0.00453 0.94233± 0.00503 0.93809± 0.00446 0.93917± 0.00537 ✓ 0.94825± 0.00399 0.94889± 0.00392 0.94778± 0.00428 0.94811± 0.00474 AC-SG3 0.92815± 0.00595 0.93158± 0.00645 0.92868± 0.00583 0.93101± 0.00604 ✓ 0.93551± 0.00546 0.93700± 0.00566 0.93289± 0.00485 0.93300± 0.00555 ViTDosovitskiy et al. (2021) No aug. ✗ 0.91820± 0.00304 0.92210± 0.00323 0.91727± 0.00321 0.92008± 0.00318 Tr.aug.(base) ✗ 0.93722± 0.00395 0.93989± 0.00402 0.93701± 0.00401 0.93817± 0.00432 SD(txt) 0.89552± 0.00392 0.89981± 0.00416 0.89461± 0.00364 0.89777± 0.00390 ✓ 0.93133± 0.00374 0.93172± 0.00345 0.93175± 0.00322 0.93156± 0.00332 SD(txt+img) 0.95095± 0.00436 0.95050± 0.00408 0.94976± 0.00452 0.94923± 0.00439 ✓ 0.95804± 0.00397 0.95812± 0.00401 0.95787± 0.00389 0.95893± 0.00397 SG3(lbl) 0.94265± 0.00401 0.94819± 0.00362 0.94678± 0.00360 0.94735± 0.00370 ✓ 0.948911± 0.00393 0.95293± 0.00323 0.94809± 0.00300 0.934784± 0.00353 AC-SG3 0.93021± 0.00613 0.93465± 0.00617 0.92999± 0.00523 0.93373± 0.00614 ✓ 0.93935± 0.00404 0.94080± 0.00430 0.93863± 0.00499 0.93901± 0.00462 Several relevant trends emerge from this analysis. Traditional data augmentation consistently outperforms training without augmentation by 1–2%, confirming its effectiveness in enhancing model generalizability. Regarding the different synthetic data augmentation techniques, highlighted in gray, we observe setups where SIF improves performance over non-SIF counterparts and surpasses traditional augmentation. The adoption of synthetic data augmentation without SIF does not necessarily yield improvements. Indeed, on both PhotoMOCI and KOCD, integrating synthetic data produced via SD(txt) and AC-SG3 without SIF (represented in the first row of each two-row block) results in either a performance degradation or marginal gains that fail to surpass traditional augmentation, regardless of the classifier used. However, by filtering images using SIF, a performance recovery is achieved when using AC-SG3 in three out of four cases (with ViT and ResNet on KOCD). This suggests that AC-SG3-based synthetic augmentation can generate high-quality images that, when selected by SIF, successfully drive performance gains. In contrast, this improvement is absent in the SD(txt) setup, where accuracy after applying SIF merely matches the results of training without augmentation. This is further validated by the composition of the filtered dataset, which consists of only a few dozen images compared to the hundreds of synthetic images generated. This indicates that SIF filters out the majority of synthetic images, effectively reverting to a dataset size comparable to the original real samples. Finally, regarding SD(txt+img) and SG3(lbl), we observe an improvement even when SIF is not applied (around 1–2%) over the baseline, suggesting that both generative techniques produce high-quality images that improve generalization. Furthermore, these improvements become even more pronounced when SIF is adopted. Specifically, we observe an additional improvement of 0.520% and 0.608% when using SD(txt+img) and SG3(lbl), respectively. ResNet50, Tr. aug. ResNet50, SD(txt+img)+SIF ViT, Tr. aug. ViT, SD(txt+img)+SIF Figure 6: Confusion matrices for PhotoMOCI comparing traditional augmentation and SD(txt+img)+SIF for ResNet50 and ViT. To complement the aggregate metrics reported in Table 7, Figure 6 reports confusion matrices on PhotoMOCI for the traditional augmentation baseline and for the best-performing synthetic generation setup, SD(txt+img)+SIF, using both ResNet50 and ViT. The confidence intervals reported in each cell correspond to the 95% confidence intervals over the 10 runs. The confusion matrices confirm the trend in Table 7: SD(txt+img)+SIF improves neoplastic classification for both ResNet50 (31 to 33 correct predictions) and ViT (30 to 31), while also improving traumatic-lesion classification for ViT (30 to 31). Thus, the performance gain does not come at the cost of neoplastic recognition. To complement the confidence intervals and assess the statistical significance of the observed improvements, we performed hypothesis tests on the same 10 runs used to compute Tables 7 and 8. First, we performed overall non-parametric repeated-measures tests to assess the statistical significance of the observed improvements when introducing SIF. Specifically, for each dataset and classifier, we applied the Friedman test χ^2 across all training setups, considering the 10 repeated runs as paired blocks, following established recommendations for non-parametric comparison of learning methods Demšar (2006); García et al. (2010); Derrac et al. (2011). Kendall’s W was reported as effect size Tomczak and Tomczak (2014). From Table 9, we observe that the Friedman test showed a significant overall effect of the training setup on accuracy for both datasets and classifiers (p<.0001p<.0001), with Kendall’s W ranging from 0.564 to 0.631, indicating a moderate-to-large effect. Table 9: Overall statistical significance analysis using Friedman tests on accuracy. Dataset Model χ^2 p Kendall’s W W KOCD ResNet 54.938 <0.0001<0.0001 0.610 KOCD ViT 53.520 <0.0001<0.0001 0.595 PhotoMOCI ResNet 56.749 <0.0001<0.0001 0.631 PhotoMOCI ViT 50.787 <0.0001<0.0001 0.564 We then performed planned paired Wilcoxon signed-rank tests, following recommendations for paired non-parametric comparisons of learning methods Demšar (2006); Benavoli et al. (2016). Two comparison families were considered: (i) each synthetic generation setup combined with SIF was compared against traditional augmentation; and (i) each SIF-filtered setup was compared against its corresponding unfiltered synthetic setup. From Table 10, we observe that, against traditional augmentation, SD(txt+img)+SIF was the most consistent setup, improving accuracy in all four cases with significant effects: KOCD ResNet (+2.38 p, p=.0371p=.0371), KOCD ViT (+2.08 p, p=.0059p=.0059), PhotoMOCI ResNet (+1.73 p, p=.0273p=.0273), and PhotoMOCI ViT (+2.35 p, p=.0195p=.0195). SG3(lbl)+SIF also improved accuracy in all four cases, but reached significance only for KOCD ResNet (+1.72 p, p=.0488p=.0488), with a borderline effect on PhotoMOCI ViT (+1.79 p, p=.0515p=.0515). Conversely, SD(txt)+SIF and AC-SG3+SIF did not consistently surpass traditional augmentation, confirming that the benefit of SIF depends on the quality and conditioning of the synthetic generation setup. Table 10: Planned Wilcoxon comparisons between synthetic+SIF setups and traditional augmentation on accuracy. Values report accuracy difference in percentage points and p-value. KOCD PhotoMOCI Synthetic img. gen. setup ResNet ViT ResNet ViT Δ Acc. p Δ Acc. p Δ Acc. p Δ Acc. p SD(txt)+SIF -1.49 0.0645 -0.59 0.3750 -2.44 0.0371 -0.84 0.4316 SD(txt+img)+SIF +2.38 0.0371 +2.08 0.0059 +1.73 0.0273 +2.35 0.0195 SG3(lbl)+SIF +1.72 0.0488 +1.17 0.0934 +1.08 0.0754 +1.79 0.0515 AC-SG3+SIF +0.44 0.6953 +0.21 0.6250 -1.53 0.0195 -0.17 0.6250 Finally, Table 11 compares each SIF-filtered setup with its corresponding unfiltered synthetic augmentation. SIF produced positive accuracy differences in all 16 planned comparisons. The largest gains were observed for SD(txt), with improvements of +3.05 p and +3.58 p on KOCD and +5.92 p and +6.31 p on PhotoMOCI, all with p≤.0039p≤.0039. For SD(txt+img), the gains were smaller (+0.48 to +0.80 p), with significant improvements for both ViT classifiers and borderline results for both ResNet classifiers (p=.0566p=.0566). SG3(lbl) and AC-SG3 also showed mostly positive but sometimes borderline evidence; significant gains were observed for SG3(lbl) on PhotoMOCI ResNet (+0.32 p, p=.0250p=.0250) and for AC-SG3 on KOCD ViT (+0.91 p, p=.0223p=.0223), PhotoMOCI ResNet (+1.84 p, p=.0195p=.0195), and PhotoMOCI ViT (+1.33 p, p=.0324p=.0324). Overall, although several cases lie close to conventional significance thresholds, the direction of the paired differences is consistently favorable to SIF. Table 11: Planned Wilcoxon comparisons between SIF-filtered and unfiltered synthetic augmentation on accuracy. Values report accuracy difference in percentage points and p-value. KOCD PhotoMOCI Synthetic img. gen. setup ResNet ViT ResNet ViT Δ Acc. p Δ Acc. p Δ Acc. p Δ Acc. p SD(txt) +3.05 0.0039 +3.58 0.0020 +5.92 0.0020 +6.31 0.0020 SD(txt+img) +0.48 0.0566 +0.71 0.0316 +0.60 0.0566 +0.80 0.0316 SG3(lbl) +0.81 0.0754 +0.63 0.0750 +0.32 0.0250 +0.57 0.0695 AC-SG3 +0.74 0.0602 +0.91 0.0223 +1.84 0.0195 +1.33 0.0324 5.4 Ablation In this section, we conduct two distinct ablation studies: (i) to evaluate the efficacy of Synthetic Image Filtering by isolating the individual contributions of Synthetic Proxy Classifier and the Synthetic Image Detector and to assess the SID threshold calibration. SIF component. The results for the individual contributions of Synthetic Proxy Classifier and the Synthetic Image Detector are presented in Figure 7(a) and 7(b) for PhotoMOCI and KOCD, respectively. For a comparative performance analysis, we established as a baseline the training on the small real dataset (blue bar). We then evaluated four distinct filtering augmentation configurations using ViT as oral lesion classifier, which demonstrated performance trends consistent with the ResNet architecture. The filtering configurations included: (i) an unfiltered dataset generated by completely omitting SIF (ochre bar); (i) a filtered dataset utilizing only the SPC component (water-green bar); (i) a filtered dataset utilizing only the SID component (orange bar); (iv) the complete SIF pipeline integrating both auxiliary classifiers (pink bar). From the figures, we observe that the SD(txt+img) and SG3(lbl) configurations exhibit consistent behavior across both PhotoMOCI and KOCD. As previously discussed, the introduction of synthetic data in these settings provides a benefit even in the absence of SIF. Nevertheless, the adoption of SIF further yields marginal accuracy improvements, with SPC contributing more than SID. This indicates that, when synthetic images are generally of good quality, selecting samples through Synthetic Proxy Classifier that contribute to learning more discriminative class boundaries is crucial for improving classifier performance. Conversely, in scenarios where the introduction of synthetic data leads to performance degradation (SD(txt) on both PhotoMOCI and KOCD, and AC-SG3 on PhotoMOCI), an opposite trend emerges. In such cases, the primary gains are provided by SID, suggesting that low realism and visually implausible synthetic images are the main source of noise responsible for the observed drop in performance, and that their removal mitigates negative effects. (a) (b) Figure 7: Ablation study evaluating the impact of SPC and SID within the SIF design on (a) PhotoMOCI and (b) KOCD using ViT. The study compares training without augmentation (blue bar) to training with integrated synthetic data: unfiltered (ochre bar), filtered using SPC only (water-green bar), SID only (orange bar), and SIF (pink bar). SID threshold calibration. To assess the sensitivity of SIF to the decision threshold used by the Synthetic Image Detector, we conducted an additional calibration experiment by varying the SID threshold τ in the interval [0,1][0,1] with a step of 0.010.01. The analysis was performed for both datasets, both classifiers (ResNet50 and ViT), and all synthetic data augmentation setups. For each threshold value, we derived the corresponding retained synthetic images and evaluated the final classification accuracy. The value τ=0τ=0 corresponds to training without synthetic images, τ=0.5τ=0.5 corresponds to the default SIF configuration, and τ=1τ=1 corresponds to the unfiltered synthetic augmentation setting. Confidence bands report the 95% confidence intervals estimated over the same 10 runs, consistent with all the other experiments where we repeated multiple runs. Figure 8 summarizes the calibration analysis and shows that the final classification performance is stable around the threshold value τ=0.5τ=0.5. Additionally, the limited overlap of confidence intervals among different synthetic data augmentation setups indicates that the relative differences among setups are preserved across the threshold sweep. In the few PhotoMOCI cases where the no-augmentation baseline is already higher than the corresponding SIF result, namely ResNet50 with SD(txt), ResNet50 with AC-SG3, and ViT with SD(txt), the maximum is attained at τ=0τ=0, consistently with the values reported in Table 7. (a) (b) Figure 8: SID threshold calibration on (a) PhotoMOCI and (b) KOCD. Each plot reports accuracy as a function of the SID threshold τ for ResNet50 (left) and ViT (right), with shaded 95% confidence intervals over 10 runs and a marker indicating the best threshold. 6 Discussion A preliminary consideration about current photographic oral cancer datasets concerns the practical relevance of binary cancer vs non-cancer (healthy) classifications for photographic screening. Although such datasets support controlled experimentation, especially for benchmarking, they do not reflect real clinical workflows when the negative class consists of healthy mucosa. Physicians would not capture images when no suspicious lesion is visible; rather, photographic assessment is typically relevant after a lesion has already been observed, and the clinical question becomes how to prioritize, characterize, or refer it. Consequently, binary datasets may overstate their utility and fail to represent the operational context in which computer-assisted tools are intended to function. In this sense, the three-class PhotoMOCI formulation, distinguishing aphthous, traumatic, and neoplastic lesions, better approximates a lesion-level triage scenario, where the model must discriminate among visible pathological findings instead of separating pathology from healthy tissue. Summarizing the experiment results, traditional data augmentation consistently provides solid performance and remains a safe and reliable baseline strategy. Regarding synthetic data augmentation strategies, the effectiveness of the proposed SIF strongly depends on the quality of the images generated during the earlier Synthetic Image Generation step of the proposed methodology. If the majority of the generated images are of low quality, SIF filters out most samples, resulting in a training setup that closely resembles training without augmentation on small real datasets. This behavior was observed in the ablation study, specifically, in the setup involving SD(txt) on the PhotoMOCI dataset. Although the original datasets are characterized by a relatively balanced class, the analysis reported in Table 6 shows that the SIF preserves this balance after filtering, without favoring any particular pathology. More generally, class imbalance, whether given by the original dataset distributions or introduced during synthetic image generation, is not a limitation of the proposed pipeline. If required, sample discarding strategies during a next step can be adopted to increase the number of retained samples for under-represented categories. Therefore, the adoption of generative models does not inherently guarantee performance improvements when synthetic samples are incorporated without further selection. The benefit of synthetic data augmentation depends primarily on two factors: (i) the architecture of the generative model and the adopted conditioning mechanism. In some cases, fine-tuning these models can be unstable or may lead to generators that fail to generalize properly, even when trained on limited datasets, as is typical for oral photographic datasets. This phenomenon is evident in the qualitative samples produced by SD(txt) and AC-SG3 in Figure 4, as well as by SD(txt) in Figure 5. (i) The nature of the data, which may be highly granular, as in the case of the PhotoMOCI dataset. Such granularity increases the difficulty of generating realistic samples that faithfully capture the distinctive visual characteristics of the target classes. Based on these considerations, we outline here some practical guidelines for the usage of the proposed methodology illustrated in Figure 2. Specifically, this methodology should not be considered as a universal plug-and-play pipeline capable of returning an optimal oral lesion classifier. Instead, given the numerous factors affecting DL classifier training (such as the number of classes, class imbalance, single-center data collection, acquisition protocol, limited sample size, or class granularity), it should be considered as a set of guidelines for carefully evaluating each stage of the pipeline before proceeding to the next one. Indeed, these aspects may introduce dataset- and acquisition-specific biases, and should be considered when interpreting the results. The proposed training pipeline presents some drawbacks compared to DL training based on traditional data augmentation techniques involving geometric and photometric transformations. These drawbacks mainly relate to increased computational requirements and additional human effort. While traditional data augmentation relies on training a single classifier, where augmentations are typically applied at runtime and are well supported by modern DL libraries, the proposed methodology presents the following practical considerations: • It requires training two additional auxiliary classifiers, SPC and SID, which implement SIF, as well as fine-tuning a generative model, which represents the most computationally demanding component. • About the development of the generative architecture, before proceeding to filtering and training the oral lesion classifier, it is essential to perform a preliminary evaluation of the generated images. This evaluation should ensure an appropriate balance between fidelity to the original image distribution (assessed using FID and KID metrics) and variability (measured through Precision, Recall, and Coverage), where generating samples that cover multiple patterns and cases facilitates the generalization of the final classifier. • The development of SIF involves the training of both SPC and SID. In this context, achieving a balance between the high performance of SPC in addressing the same downstream classification task on synthetic images and the near-random performance of SID in failing to distinguish real from synthetic images is crucial. As shown in Table 5, SD(txt) clearly exhibits difficulties in generating high-quality synthetic images. The same considerations are less intuitive if we compare SD(txt+img) and SG3(lbl) on PhotoMOCI. The first one gets higher SPC accuracy, while the second one achieves a better (lower) accuracy value on SID. • The methodology is more suitable for a two-phase training strategy, in which the Filtered Dataset is precomputed after the Synthetic Image Generation phase and the application of SIF. Generating synthetic samples and evaluating them with SIF on-the-fly during training would be particularly challenging, especially for diffusion-based models, whose inference process involves multiple diffusion steps. As reported in Table 4, generating the complete SD(txt+img)-based image pool requires around 28 and 41 min for PhotoMOCI and KOCD, respectively, whereas ResNet50-based SIF filtering requires only 10.1 ms per image; thus, once generation is completed offline, filtering adds minimal overhead to dataset precomputation. • Despite recent advances aimed at stabilizing training through regularization techniques and reducing computational costs using approaches such as LoRA Yang et al. (2024), training a generative model remains challenging, particularly in limited-data scenarios. This often results in a time-consuming process, where substantial effort is devoted to hyperparameter tuning before obtaining a satisfactory generative model. Under the assumption that sufficient computational resources are available and that multiple training trials, or preferably a systematic hyperparameter optimization procedure, can be conducted, the proposed methodology represents the most effective strategy among those evaluated in our benchmark for improving classification performance under the considered experimental setting, as indicated in Tables 7 and 8. However, when these conditions cannot be met, traditional data augmentation remains the preferred solution, as it is well supported by existing DL libraries, lightweight in terms of computational overhead, and can be efficiently applied at runtime during model training. 7 Conclusion This work addresses the challenges of data scarcity and investigates the synthetic data augmentation in photographic oral cancer. We introduced PhotoMOCI, a novel dataset designed to support multiple diagnostic tasks in the oral cancer domain, providing a publicly available resource for open research and benchmarking. We also conduct a comprehensive evaluation of data augmentation strategies, comparing traditional transformation-based methods with generative approaches based on GANs and Diffusion Models. Our findings indicate that the direct use of synthetic images does not consistently improve classification performance and may even be counterproductive, emphasizing that synthetic data quality is critical in this setting. To address this limitation, we proposed the Synthetic Image Filter, which selects synthetic samples based on machine-assessed visual realism and class representativeness using two auxiliary models: Synthetic Proxy Classifier and Synthetic Image Detector. Although qualitative inspection shows visual analogies between some retained synthetic images and real oral lesions, the term realism in this work refers specifically to machine indistinguishability from the real-image distribution, rather than to clinical realism or diagnostic plausibility. Experimental results show that this selective strategy effectively removes low-utility images and leads to more stable and improved downstream performance. Overall, this study highlights the importance of careful dataset design and controlled use of generative augmentation in photographic oral cancer imaging, offering a comprehensive benchmark for future research on reliable data augmentation pipelines. At the same time, the reported improvements should be regarded as dataset-specific experimental findings. The single-center origin of PhotoMOCI and the absence of an human expert assessment of the SIF-selected images represent a limitation; external and clinical validation would be strongly recommended before assessing its applicability in real-world screening workflows. 7.1 Future work The proposed pipeline enables a multi-level evaluation of synthetic images and their impact when used as data augmentation for downstream tasks such as image classification. Specifically, (i) assessing generative image metrics at the Synthetic Image Generation stage, (i) employing auxiliary classifiers assess image realism and class representativeness in the Synthetic Image Filter, while (i) the final evaluation considers the effect of selected synthetic data on downstream performance. Given the importance of trustworthiness when deploying AI in high-risk domains such as healthcare, model transparency and alignment with human judgment are crucial Longo et al. (2024); Parola et al. (2026); Rong et al. (2023). References Alzubaidi et al. (2023) L. Alzubaidi, J. Bai, A. Al-Sabaawi, J. Santamaría, A. Albahri, B. Al-dabbagh, M. Fadhel, M. Manoufali, J. Zhang, A. Al-Timemy, Y. Duan, A. Abdullah, L. Farhan, Y. Lu, A. Gupta, F. Albu, A. Abbosh, and Y. Gu A survey on deep learning tools dealing with data scarcity: definitions, challenges, solutions, tips, and applications. Journal of Big Data 10, p. . External Links: Document Cited by: §1, §1, §2.1. Arjovsky et al. (2017) M. Arjovsky, S. Chintala, and L. Bottou Wasserstein generative adversarial networks. In Proceedings of the 34th International Conference on Machine Learning (ICML), Vol. 70, p. 214–223. External Links: Link Cited by: §2.1. Benavoli et al. (2016) A. Benavoli, G. Corani, and F. Mangili Should we really use post-hoc tests based on mean-ranks?. Journal of Machine Learning Research 17 (5), p. 1–10. Cited by: §5.3. Bińkowski et al. (2018) M. Bińkowski, D. J. Sutherland, M. Arbel, and A. Gretton Demystifying mmd gans. arXiv preprint arXiv:1801.01401. Cited by: §5.1. Bourou et al. (2024) A. Bourou, V. Mezger, and A. Genovesio GANs conditioning methods: a survey. arXiv preprint arXiv:2408.15640. Cited by: §2.2, §2.2. Bramati et al. (2021) C. Bramati, S. Abati, S. Bondi, A. Lissoni, G. Arrigoni, F. Filipello, and M. Trimarchi Early diagnosis of oral squamous cell carcinoma may ensure better prognosis: a case series. Clinical Case Reports 9 (10). Cited by: §1. Brooks (2019) J. Brooks COCO Annotator. Note: https://github.com/jsbroks/coco-annotator/ Cited by: §3. Cimino et al. (2025) M. G.C.A. Cimino, G. Campisi, F. A. Galatolo, P. Neri, P. Tozzo, M. Parola, G. La Mantia, and O. Di Fede Explainable screening of oral cancer via deep learning and case-based reasoning. Smart Health 35, p. 100538. External Links: ISSN 2352-6483, Document, Link Cited by: §1, §3. Demšar (2006) J. Demšar Statistical comparisons of classifiers over multiple data sets. Journal of Machine Learning Research 7, p. 1–30. Cited by: §5.3, §5.3. Deng et al. (2009) J. Deng, W. Dong, R. Socher, L. Li, K. Li, and L. Fei-Fei ImageNet: a large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, Vol. , p. 248–255. External Links: Document Cited by: §5.3, §5.3. Derrac et al. (2011) J. Derrac, S. García, D. Molina, and F. Herrera A practical tutorial on the use of nonparametric statistical tests as a methodology for comparing evolutionary and swarm intelligence algorithms. Swarm and Evolutionary Computation 1 (1), p. 3–18. External Links: Document Cited by: §5.3. Dhariwal and Nichol (2021) P. Dhariwal and A. Nichol Diffusion models beat gans on image synthesis. In Advances in Neural Information Processing Systems, M. Ranzato, A. Beygelzimer, Y. Dauphin, P.S. Liang, and J. W. Vaughan (Eds.), Vol. 34, p. 8780–8794. External Links: Link Cited by: §2.1. Di Fede et al. (2024) O. Di Fede, G. La Mantia, M. Parola, L. Maniscalco, D. Matranga, P. Tozzo, G. Campisi, and M. G. Cimino Automated detection of oral malignant lesions using deep learning: scoping review and meta-analysis. Oral Diseases. Cited by: §1, §1. Dos Santos et al. (2023) D. F. D. Dos Santos, P. R. de Faria, B. A. N. Travençolo, and M. Z. do Nascimento Influence of data augmentation strategies on the segmentation of oral histological images using fully convolutional neural networks. Journal of Digital Imaging 36 (4). External Links: Document Cited by: §2.1. Dosovitskiy et al. (2021) A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby An image is worth 16x16 words: transformers for image recognition at scale. In International Conference on Learning Representations, Cited by: §5.3, Table 7, Table 8. Frid-Adar et al. (2018) M. Frid-Adar, I. Diamant, E. Klang, M. Amitai, J. Goldberger, and H. Greenspan GAN-based synthetic medical image augmentation for increased cnn performance in liver lesion classification. Neurocomputing 321, p. 321–331. External Links: ISSN 0925-2312, Link, Document Cited by: §2.1. Fujimoto et al. (2022) T. Fujimoto, E. Fukuzawa, S. Tatehara, K. Satomura, and J. Ohya Automatic diagnosis of early-stage oral cancer and precancerous lesions from ala-pdd images using gan and cnn. In 2022 44th Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC), p. 2161–2164. External Links: Document Cited by: §2.1. García et al. (2010) S. García, A. Fernández, J. Luengo, and F. Herrera Advanced nonparametric tests for multiple comparisons in the design of experiments in computational intelligence and data mining: experimental analysis of power. Information Sciences 180 (10), p. 2044–2064. External Links: Document Cited by: §5.3. Goodfellow et al. (2014) I. J. Goodfellow et al. Generative adversarial networks. arXiv preprint arXiv:1406.2661. External Links: Link Cited by: §2.1, §2. He et al. (2016) K. He, X. Zhang, S. Ren, and J. Sun Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, p. 770–778. Cited by: §5.3, Table 7, Table 8. Heo et al. (2022) J. Heo, J. H. Lim, H. R. Lee, J. Y. Jang, Y. S. Shin, D. Kim, J. Y. Lim, Y. M. Park, Y. W. Koh, S. Ahn, et al. Deep learning model for tongue cancer diagnosis using endoscopic images. Scientific reports 12 (1), p. 6281. Cited by: §3. Heusel et al. (2017) M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems 30. Cited by: §5.1. Hinton and Salakhutdinov (2006) G. E. Hinton and R. R. Salakhutdinov Reducing the dimensionality of data with neural networks. Science 313 (5786), p. . External Links: Document, Link Cited by: §2.1. Ho et al. (2020) J. Ho, A. Jain, and P. Abbeel Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 33, p. 6840–6851. External Links: Link Cited by: §2.1, §2.2. Hu et al. (2021) E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, and W. Chen LoRA: low-rank adaptation of large language models. CoRR abs/2106.09685. External Links: Link, 2106.09685 Cited by: §2.1. Islam et al. (2023) Md. M. Islam, K. M. R. Alam, J. Uddin, I. Ashraf, and M. A. Samad Benign and malignant oral lesion image classification using fine-tuned transfer learning techniques. Diagnostics 13 (21). External Links: Link, ISSN 2075-4418, Document Cited by: Table 1. Jalali et al. (2023) M. Jalali, C. T. Li, and F. Farnia An information-theoretic evaluation of generative models in learning multi-modal distributions. In Advances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine (Eds.), Vol. 36, p. 9931–9943. External Links: Link Cited by: §5.1. Johnson and Khoshgoftaar (2019) J. M. Johnson and T. M. Khoshgoftaar Survey on deep learning with class imbalance. Journal of Big Data 6 (1), p. 27. External Links: Document, Link Cited by: §2.1, §2. Jubair et al. (2022) F. Jubair, O. Al-karadsheh, D. Malamos, S. Al Mahdi, Y. Saad, and Y. Hassona A novel lightweight deep convolutional neural network for early detection of oral cancer. Oral Diseases 28 (4), p. 1123–1130. Cited by: §3. Karras et al. (2021) T. Karras, M. Aittala, S. Laine, E. Härkönen, J. Hellsten, J. Lehtinen, and T. Aila Alias-free generative adversarial networks. Advances in neural information processing systems 34, p. 852–863. Cited by: §4.1. Karras et al. (2019) T. Karras, S. Laine, and T. Aila A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, Cited by: §2.1. Karras et al. (2020) T. Karras et al. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 8110–8119. Cited by: §2.1. Kazemzadeh et al. (2024) S. Kazemzadeh et al. Addressing class imbalance with latent diffusion-based data augmentation for disease classification in pediatric chest x-rays. Scientific Reports 14 (1), p. 7908. External Links: Link Cited by: §2.1, §2.1, §2.2. Keane and Kenny (2019) M. T. Keane and E. M. Kenny How case-based reasoning explains neural networks: a theoretical analysis of xai using post-hoc explanation-by-example from a survey of ann-cbr twin-systems. In International conference on case-based reasoning, p. 155–171. Cited by: §3. Kenny and Keane (2019) E. M. Kenny and M. T. Keane Twin-systems to explain artificial neural networks using case-based reasoning: comparative tests of feature-weighting methods in ann-cbr twins for xai. In Kraus, S.(ed.). Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, Cited by: §3. Kim and Kim (2020) Y. Kim and J. H. Kim Increasing incidence and improving survival of oral tongue squamous cell carcinoma. Scientific reports 10 (1), p. 7877. Cited by: §1. Kingma and Welling (2013) D. P. Kingma and M. Welling Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114. External Links: Link Cited by: §2.1. Kingma (2014) D. P. Kingma Adam: a method for stochastic optimization. arXiv preprint arXiv:1412.6980. Cited by: §5.2, §5.3. Kouketsu et al. (2024) A. Kouketsu, C. Doi, H. Tanaka, T. Araki, R. Nakayama, T. Toyooka, S. Hiyama, M. Iikubo, K. Osaka, K. Sasaki, et al. Detection of oral cancer and oral potentially malignant disorders using artificial intelligence-based image analysis. Head & Neck. Cited by: §3. Krizhevsky et al. (2012) A. Krizhevsky, I. Sutskever, and G. E. Hinton ImageNet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems (NeurIPS), Vol. 25. External Links: Link Cited by: §2.1, §2.1. Kynkänniemi et al. (2019) T. Kynkänniemi, T. Karras, S. Laine, J. Lehtinen, and T. Aila Improved precision and recall metric for assessing generative models. Advances in neural information processing systems 32. Cited by: §5.1. Lee et al. (2021) C. H. Lee, Z. Zhang, and X. Zhao A survey of smart healthcare for the elderly based on user requirements and supply accessibility. In 5th International Conference on Crowd Science and Engineering, Cited by: §1. Lee et al. (2023) S. Lee, I. Kwon, Y. Son, J. Kim, D. Kwon, B. Kim, J. Lee, and H. Kim Early tongue cancer detection in photographs using a pretrained convolutional neural network. Cited by: §3. Lin et al. (2021) H. Lin, H. Chen, L. Weng, J. Shao, and J. Lin Automatic detection of oral cancer in smartphone-based images using deep learning for early diagnosis. Journal of Biomedical Optics 26 (8), p. 086007–086007. Cited by: §3. Lin et al. (2014) T. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick Microsoft coco: common objects in context. In European conference on computer vision, p. 740–755. Cited by: §5.3. Liu et al. (2020) S. Liu, T. Wang, D. Bau, J. Zhu, and A. Torralba Diverse image generation via self-conditioned gans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 14286–14295. Cited by: §2.2. Longo et al. (2024) L. Longo, M. Brcic, F. Cabitza, J. Choi, R. Confalonieri, J. Del Ser, R. Guidotti, Y. Hayashi, F. Herrera, A. Holzinger, et al. Explainable artificial intelligence (xai) 2.0: a manifesto of open challenges and interdisciplinary research directions. Information Fusion 106, p. 102301. Cited by: §7.1. Mantegna et al. (2025) M. Mantegna, L. Tronchin, M. Tortora, and P. Soda Benchmarking gan-based vs classical data augmentation on biomedical images. In Pattern Recognition. ICPR 2024 International Workshops and Challenges, S. Palaiahnakote, S. Schuckers, J. Ogier, P. Bhattacharya, U. Pal, and S. Bhattacharya (Eds.), Cham, p. 92–104. External Links: ISBN 978-3-031-87660-8 Cited by: §1, §2.1. Mirza and Osindero (2014) M. Mirza and S. Osindero Conditional generative adversarial nets. External Links: 1411.1784, Link Cited by: §2.1. Naeem et al. (2020) M. F. Naeem, S. J. Oh, Y. Uh, Y. Choi, and J. Yoo Reliable fidelity and diversity metrics for generative models. In International conference on machine learning, p. 7176–7185. Cited by: §5.1. Neff et al. (2017) T. Neff, C. Payer, D. Štern, and M. Urschler Generative adversarial network based synthesis for supervised medical image segmentation. In Proceedings of the OAGM & ARW Joint Workshop on Vision, Automation and Robotics, p. 140–145. External Links: Document Cited by: §2.1. Odena et al. (2017) A. Odena, C. Olah, and J. Shlens Conditional image synthesis with auxiliary classifier gans. In International conference on machine learning, p. 2642–2651. Cited by: §2.2. Oya et al. (2023) K. Oya, K. Kokomoto, K. Nozaki, and S. Toyosawa Oral squamous cell carcinoma diagnosis in digitized histological images using convolutional neural network. Journal of Dental Sciences 18 (1). External Links: ISSN 1991-7902, Document, Link Cited by: §2.1. Parola et al. (2026) M. Parola, A. L. Alfeo, and M. G. Cimino Human-centered xai via a concept-informed prompt-based validation framework for saliency maps [ciprova]. Image and Vision Computing, p. 105920. Cited by: §7.1. Parola et al. (2025a) M. Parola, M. G.C.A. Cimino, I. Cantini, G. L. Mantia, G. Campisi, and O. Di Fede Oral cancer recognition on photographic images via deep learning semantic segmentation. In 2025 IEEE Symposium on Computational Intelligence in Health and Medicine Companion (CIHM Companion), Vol. , p. 1–5. External Links: Document Cited by: §3. Parola et al. (2024) M. Parola, F. Galatolo, G. La Mantia, M. G.C.A. Cimino, G. Campisi, and O. Di Fede Towards explainable oral cancer recognition: screening on imperfect images via informed deep learning and case-based reasoning. Computerized Medical Imaging and Graphics 117, p. 102433. External Links: ISSN 0895-6111, Document, Link Cited by: §1, §3. Parola et al. (2025b) M. Parola, S. Landi, and F. Nocella GitHub synthetic oral cancer experiment code repository, https://github.com/marcoparola/oral3. External Links: Link Cited by: §5, §5. Parola et al. (2023) M. Parola, G. L. Mantia, F. Galatolo, M. G.C.A. Cimino, G. Campisi, and O. Di Fede Image-based screening of oral cancer via deep ensemble architecture. In 2023 IEEE Symposium Series on Computational Intelligence (SSCI), Vol. , p. 1572–1578. External Links: Document Cited by: §3. Parola (2025) M. Parola PhotoMOCI dataset. Kaggle. Note: KaggleAccessed 14/03/2025 External Links: Link, Document Cited by: §3, §5, Abstract. Pinaya et al. (2024) W. H. L. Pinaya et al. Adapted generative latent diffusion models for accurate pathological analysis in chest x-ray images. Medical Image Analysis. External Links: Link Cited by: §2.1. Qasim et al. (2021) A. B. Qasim, I. Ezhov, S. Shit, O. Schoppe, J. C. Paetzold, A. Sekuboyina, F. Kofler, J. Lipkova, H. Li, and B. Menze Red-gan: attacking class imbalance via conditioned generation. yet another perspective on medical image synthesis for skin lesion dermoscopy and brain tumor mri. External Links: 2004.10734, Link Cited by: §2.1. Qi et al. (2020) C. Qi, J. Chen, G. Xu, Z. Xu, T. Lukasiewicz, and Y. Liu SAG-gan: semi-supervised attention-guided gans for data augmentation on medical images. External Links: 2011.07534, Link Cited by: §2.1. Rabinovici-Cohen et al. (2024) S. Rabinovici-Cohen, N. Fridman, M. Weinbaum, E. Melul, E. Hexter, M. Rosen-Zvi, Y. Aizenberg, and D. Porat Ben Amy From pixels to diagnosis: algorithmic analysis of clinical oral photos for early detection of oral squamous cell carcinoma. Cancers 16 (5), p. 1019. Cited by: §3. Radford et al. (2021) A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, p. 8748–8763. Cited by: 1st item. Radford et al. (2015) A. Radford, L. Metz, and S. Chintala Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434. External Links: Link Cited by: §2.1. Rajpurkar et al. (2022) P. Rajpurkar, E. Chen, O. Banerjee, and E. J. Topol AI in health and medicine. Nature medicine 28 (1), p. 31–38. Cited by: §1. Rashid (2024) M. Z. Rashid Oral cancer images, kaggle. Note: https://w.kaggle.com/datasets/zaidpy/oral-cancer-datasetAccessed: 14/03/2025 External Links: Link Cited by: Table 1, Table 5, §5. Ribeiro et al. (2022) M. F. A. Ribeiro, M. C. M. Oliveira, A. C. Leite, F. F. B. Bruzinga, P. A. Mendes, S. d. M. C. Grossmann, V. E. de Araújo, and G. R. Souto Assessment of screening programs as a strategy for early detection of oral cancer: a systematic review. Oral Oncology 130. Cited by: §1. Rofena et al. (2025a) A. Rofena, A. Manchia, C. L. Piccolo, B. B. Zobel, P. Soda, and V. Guarrasi Lesion-aware generative artificial intelligence for virtual contrast-enhanced mammography in breast cancer. arXiv preprint arXiv:2505.03018. Cited by: §2.1. Rofena et al. (2025b) A. Rofena, C. L. Piccolo, B. B. Zobel, P. Soda, and V. Guarrasi Augmented intelligence for multimodal virtual biopsy in breast cancer using generative artificial intelligence. preprint arXiv:2501.19176. Cited by: §2.1. Rombach et al. (2022) R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 10684–10695. External Links: Link Cited by: §2.1, §2.2, §4.1. Rong et al. (2023) Y. Rong, T. Leemann, T. Nguyen, L. Fiedler, P. Qian, V. Unhelkar, T. Seidel, G. Kasneci, and E. Kasneci Towards human-centered explainable ai: a survey of user studies for model explanations. IEEE transactions on pattern analysis and machine intelligence 46 (4), p. 2104–2122. Cited by: §7.1. Ronneberger et al. (2015) O. Ronneberger, P. Fischer, and T. Brox U-net: convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, N. Navab, J. Hornegger, W. M. Wells, and A. F. Frangi (Eds.), Cham, p. 234–241. External Links: ISBN 978-3-319-24574-4 Cited by: §2.1, §2.1, §2.1. sagari vijay (2021) sagari vijay Roboflow oral cancer data dataset. Open Source Dataset, Roboflow. Note: https://universe.roboflow.com/sagari-vijay/oral-cancer-datavisited on 2024-11-13 External Links: Link Cited by: Table 1. Sagers et al. (2023) L. W. Sagers, J. A. Diao, L. Melas-Kyriazi, M. Groh, P. Rajpurkar, A. S. Adamson, V. Rotemberg, R. Daneshjou, and A. K. Manrai Augmenting medical image classifiers with synthetic data from latent diffusion models. External Links: 2308.12453, Link Cited by: §2.1. Shin et al. (2016) H. Shin, H. R. Roth, M. Gao, L. Lu, Z. Xu, I. Nogues, J. Yao, D. Mollura, and R. M. Summers Deep convolutional neural networks for computer-aided detection: cnn architectures, dataset characteristics and transfer learning. IEEE Transactions on Medical Imaging 35 (5), p. 1285–1298. External Links: Document Cited by: §2.1. Shivam and Prakrut (2020) B. Shivam and S. Prakrut Oral cancer (lips and tongue) images, kaggle. Note: kaggle.com/datasets/shivam17299/oral-cancer-lips-and-tongue-images External Links: Link Cited by: Table 1. Shorten and Khoshgoftaar (2019) C. Shorten and T. M. Khoshgoftaar A survey on image data augmentation for deep learning. Journal of big data 6 (1), p. . Cited by: §2.1, §2.1. Sohl-Dickstein et al. (2015) J. Sohl-Dickstein, E. A. Weiss, N. Maheswaranathan, and S. Ganguli Deep unsupervised learning using nonequilibrium thermodynamics. In Proceedings of the 32nd International Conference on Machine Learning (ICML), Vol. 37, p. 2256–2265. External Links: Link Cited by: §2.1. Tang et al. (2023) L. Tang et al. Deep learning approaches for data augmentation in medical imaging: a review. Journal of Imaging 9 (4). External Links: Link Cited by: §1, §2.1, §2.1, §2.1, §2.1, §2.1. Tomczak and Tomczak (2014) M. Tomczak and E. Tomczak The need to report effect size estimates revisited. an overview of some recommended measures of effect size. Trends in Sport Sciences 21 (1), p. 19–25. Cited by: §5.3. Uliana and Krohling (2025) J. J. Uliana and R. A. Krohling Diffusion models applied to skin and oral cancer classification. arXiv preprint arXiv:2504.00026. Cited by: §2.1. Vinayahalingam et al. (2024) S. Vinayahalingam, N. van Nistelrooij, R. Rothweiler, A. Tel, T. Verhoeven, D. Tröltzsch, M. Kesting, S. Bergé, T. Xi, M. Heiland, et al. Advancements in diagnosing oral potentially malignant disorders: leveraging vision transformers for multi-class detection. Clinical Oral Investigations 28 (7), p. 1–8. Cited by: §3. Wael Alosaimi (2022) M. I. U. Wael Alosaimi Efficient data augmentation techniques for improved classification in limited data set of oral squamous cell carcinoma. Computer Modeling in Engineering & Sciences 131 (3), p. 1387–1401. External Links: Link, ISSN 1526-1506, Document Cited by: §2.1, §2.1. Waheed et al. (2020) A. Waheed, M. Goyal, D. Gupta, A. Khanna, F. Al-Turjman, and P. R. Pinheiro CovidGAN: data augmentation using auxiliary classifier gan for improved covid-19 detection. IEEE Access 8. External Links: ISSN 2169-3536, Link, Document Cited by: §2.1. Warin et al. (2022) K. Warin, W. Limprasert, S. Suebnukarn, S. Jinaporntham, P. Jantana, and S. Vicharueang AI-based analysis of oral lesions using novel deep convolutional neural networks for early detection of oral cancer. Plos one 17 (8), p. e0273508. Cited by: §3. Welikala et al. (2021) R. A. Welikala, P. Remagnino, J. H. Lim, C. S. Chan, S. Rajendran, T. G. Kallarakkal, R. B. Zain, R. D. Jayasinghe, J. Rimal, A. R. Kerr, et al. Clinically guided trainable soft attention for early detection of oral cancer. In Computer Analysis of Images and Patterns: 19th International Conference, CAIP 2021, Virtual Event, September 28–30, 2021, Proceedings, Part I 19, p. 226–236. Cited by: §3. Whang et al. (2022) S. E. Whang, Y. Roh, H. Song, and J. Lee Data collection and quality challenges in deep learning: a data-centric ai perspective. External Links: 2112.06409, Link Cited by: §1, §2.1. Yang et al. (2024) M. Yang, J. Chen, Y. Zhang, J. Liu, J. Zhang, Q. Ma, H. Verma, Q. Zhang, M. Zhou, I. King, et al. Low-rank adaptation for foundation models: a comprehensive review. preprint arXiv:2501.00365. Cited by: §2.1, 5th item. Zhan et al. (2024) Z. Zhan, D. Chen, J. Mei, Z. Zhao, J. Chen, C. Chen, S. Lyu, and C. Wang Conditional image synthesis with diffusion models: a survey. arXiv preprint arXiv:2409.19365. Cited by: §2.2. Zhang and Qie (2023) H. Zhang and Y. Qie Applying deep learning to medical imaging: a review. Applied Sciences 13 (18). External Links: Link, ISSN 2076-3417, Document Cited by: §2. Zhou et al. (2022) J. Zhou, W. Wang, C. Zhang, Y. Ling, X. Hong, Q. Su, W. Li, Z. Mao, B. Cheng, C. Tan, et al. Ru (i)-modified tio2 nanoparticles for hypoxia-adaptive photo-immunotherapy of oral squamous cell carcinoma. Biomaterials 289, p. 121757. Cited by: §1.