Paper deep dive

Simultaneous Dual-View Mammogram Synthesis Using Denoising Diffusion Probabilistic Models

Jorge Alberto Garza-Abdala, Gerardo A. Fumagal-González, Eduardo de Avila-Armenta, Sadam Hussain, Jasiel H. Toscano-Martínezb, Diana S. M. Rosales Gurmendi, Alma A. Pedro-Pérez, Jose G. Tamez-Pena

Year: 2026Venue: arXiv preprintArea: cs.CVType: PreprintEmbeddings: 18

Intelligence

Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 96%

Last extracted: 4/10/2026, 3:01:28 AM

Summary

The paper introduces a novel method for simultaneous dual-view mammogram synthesis using a three-channel denoising diffusion probabilistic model (DDPM). By encoding craniocaudal (CC) and mediolateral oblique (MLO) views along with their absolute pixel-wise difference into an RGB image, the model learns to preserve anatomical consistency between projections. The approach was fine-tuned on a private dataset of 22,040 dual-view mammograms and demonstrated high anatomical coherence, offering a potential solution for dataset augmentation in breast cancer screening.

Entities (5)

Craniocaudal view · medical-imaging-view · 100%Denoising Diffusion Probabilistic Model · algorithm · 100%Mediolateral oblique view · medical-imaging-view · 100%Hugging Face · organization · 98%TecSalud · organization · 95%

Relation Signals (3)

TecSalud → provided → Private screening dataset

confidence 95% · This study utilized a private dataset from the TecSalud health system

Denoising Diffusion Probabilistic Model → synthesizes → Dual-view mammograms

confidence 95% · we propose a three-channel denoising diffusion probabilistic model (DDPM) capable of simultaneously generating CC and MLO views

Denoising Diffusion Probabilistic Model → uses → RGB-based encoding

confidence 90% · To enable the generation of consistent dual-view mammograms, we encoded paired views into a single RGB image.

Cypher Suggestions (0)

No Cypher suggestions yet.

Abstract

Abstract:Breast cancer screening relies heavily on mammography, where the craniocaudal (CC) and mediolateral oblique (MLO) views provide complementary information for diagnosis. However, many datasets lack complete paired views, limiting the development of algorithms that depend on cross-view consistency. To address this gap, we propose a three-channel denoising diffusion probabilistic model capable of simultaneously generating CC and MLO views of a single breast. In this configuration, the two mammographic views are stored in separate channels, while a third channel encodes their absolute difference to guide the model toward learning coherent anatomical relationships between projections. A pretrained DDPM from Hugging Face was fine-tuned on a private screening dataset and used to synthesize dual-view pairs. Evaluation included geometric consistency via automated breast mask segmentation and distributional comparison with real images, along with qualitative inspection of cross-view alignment. The results show that the difference-based encoding helps preserve the global breast structure across views, producing synthetic CC-MLO pairs that resemble real acquisitions. This work demonstrates the feasibility of simultaneous dual-view mammogram synthesis using a difference-guided DDPM, highlighting its potential for dataset augmentation and future cross-view-aware AI applications in breast imaging.

PDF

Open source PDF →Open local PDF →

Trouble viewing inline? Open PDF directly →

Full Text

17,468 characters extracted from source content.

Expand or collapse full text

Further author information: (Send correspondence to Garza-Abdala) E-mail: a01376671@tec.mx Simultaneous Dual-View Mammogram Synthesis Using Denoising Diffusion Probabilistic Models Jorge Alberto Garza-Abdala Tecnologico de Monterrey, School of Engineering and Sciences, Monterrey, Mexico Gerardo A. Fumagal-González Tecnologico de Monterrey, School of Engineering and Sciences, Monterrey, Mexico Eduardo de Avila-Armenta Tecnologico de Monterrey, School of Engineering and Sciences, Monterrey, Mexico Sadam Hussain Tecnologico de Monterrey, School of Engineering and Sciences, Monterrey, Mexico Jasiel H. Toscano-Martínez School of Engineering, Pontificia Universidad Católica de Chile, Santiago, Chile Diana S. M. Rosales Gurmendi Tecnologico de Monterrey, School of Engineering and Sciences, Monterrey, Mexico Alma A. Pedro-Pérez School of Engineering, Pontificia Universidad Católica de Chile, Santiago, Chile Jose G. Tamez-Pena Tecnologico de Monterrey, School of Medicine and Health Sciences, Monterrey, Mexico keywords: Mammography, Breast cancer, Dual-view synthesis, Diffusion models Breast cancer screening relies heavily on mammography, where the craniocaudal (C) and mediolateral oblique (MLO) views provide complementary information for diagnosis. However, many datasets lack complete paired views, limiting the development of algorithms that depend on cross-view consistency. To address this gap, we propose a three-channel denoising diffusion probabilistic model (DDPM) capable of simultaneously generating C and MLO views of a single breast. In this configuration, the two mammographic views are stored in separate channels, while a third channel encodes their absolute difference to guide the model toward learning coherent anatomical relationships between projections. A pretrained DDPM from Hugging Face was fine-tuned on a private screening dataset and used to synthesize dual-view pairs. Evaluation included geometric consistency via automated breast mask segmentation and distributional comparison with real images, along with qualitative inspection of cross-view alignment. The results show that the difference-based encoding helps preserve the global breast structure across views, producing synthetic C–MLO pairs that resemble real acquisitions. This work demonstrates the feasibility of simultaneous dual-view mammogram synthesis using a difference-guided DDPM, highlighting its potential for dataset augmentation and future cross-view–aware AI applications in breast imaging. 1 INTRODUCTION Breast cancer remains a major global health for women, with an estimated 2.3 million new cases and 670,000 deaths reported in 2022 [12]. Mammography is currently the gold standard for breast cancer screening and diagnosis due to its proven effectiveness in both early and late stages of detection[2]. Recent advances in machine learning and deep learning have significantly improved the analysis of mammographic images, enhancing the accuracy of breast cancer detection [6]. In particular, dual-view models—those that incorporate both the craniocaudal (C) and mediolateral oblique (MLO) views—have shown improved diagnostic performance over single-view approaches by leveraging complementary anatomical information [1, 11]. However, these models require large, diverse datasets to achieve robust performance, and many publicly available mammography datasets lack complete dual-view pairs[10]. To mitigate this limitation, generative approaches such as generative adversarial networks (GANs) and diffusion models have been employed to synthesize mammographic images. These efforts typically focus on generating single-view images or inferring a missing complementary view [7, 13]. In this work, we introduce a novel method for simultaneous dual-view mammogram synthesis using denoising diffusion probabilistic models (DDPMs)[4]. Our approach encodes paired views into a single RGB image, where each channel represents distinct structural information: the C view, the MLO view, and their absolute difference. This encoding allows the model to learn cross-view anatomical consistency during training. To our knowledge, this is the first approach to jointly generate both C and MLO views in a unified framework, aiming to preserve structural coherence between projections and better reflect the characteristics of real patient data. 2 METHODS 2.1 Data description and preprocessing This study utilized a private dataset from the TecSalud health system, comprising 42,454 unique mammogram studies acquired between 2014 and 2019. The dataset includes standard views for each breast—C and MLO—and was approved by the institutional ethics committee (Protocol No. P000542-MIRAI-MODIFICADO-CEIC-CR002). To ensure consistency and reduce variability, we included only the first study per patient, restricted the sample to non-implant cases, and focused on patients with BI-RADS-2 findings. This resulted in a total of 22,040 RGB dual-view mammograms from 11,020 patients, with both C and MLO views available per case. Image preprocessing involved normalizing pixel intensities to the [0, 1] range. Left breast views (C and MLO) were horizontally flipped to match the orientation of the right breast, ensuring directional consistency across the dataset. To address inter-image intensity variations, we applied histogram matching as described in previous work [3]. Finally, all images were resized to 256×256 pixels and stored in 16-bit unsigned integer (uint16) format. 2.2 RGB-based mammograms To enable the generation of consistent dual-view mammograms, we encoded paired views into a single RGB image. Specifically, the red channel corresponds to the C view, the green channel to the MLO view, and the blue channel represents the absolute pixel-wise difference between the C and MLO images. This encoding strategy allows the model to simultaneously learn view-specific features and their shared anatomical structures. As illustrated in Figure 1, the resulting RGB image serves as the input to the DDPM, enabling the joint synthesis of both views within a unified framework. Figure 1: Real mammographic views. From left to right: C; MLO; |CC−MLO| C-MLO ; The three previous views were encoded into an RGB image, where Red=C, Green=MLO, and Blue=|CC−MLO| C-MLO . 2.3 Implementation details All experiments were conducted using a DDPM on an NVIDIA Tesla A100 GPU. We initialized our model with the publicly available pretrained weights of ddpm-celebahq-256, developed by Google and hosted in the Hugging Face repository [4]. To remove residual noise from the generated images, we applied a post-processing step based on percentile normalization. Specifically, the pixel intensities were clipped at the 0.5th and 99.5th percentiles, effectively removing the lowest and highest 0.5% of values to enhance image quality. The model was fine-tuned using the following hyperparameters: Parameters: - Learning Rate: 1e-5 - Batch Size: 16 - Epochs: 100 - Scheduler: LambdaLR - Optimizer:AdamW - Betas: (0.9, 0.999) - Loss Function: Mean Squared Error 2.4 Evaluation To evaluate the consistency between synthetic C and MLO views, we generated 500 synthetic RGB images and conducted a visual inspection test, where a non-expert assessed size, alignment, and artifacts to determine if views appeared from the same patient. Additionally, we performed a quantitative analysis using binary masks obtained via Otsu’s thresholding, calculating Intersection Over Union (IoU) and Dice Similarity Coefficient (DSC) between views. The same metrics were computed on 2500 real dual-view mammograms from the TecSalud dataset to compare synthetic and real image consistency. 3 RESULTS Figure 2 presents an example of a synthetic RGB mammogram alongside the individual grayscale views extracted from each channel. The red and green channels correspond to the C and MLO views, respectively, while the blue channel captures the absolute difference between them. As observed, the model successfully maintains anatomical consistency between the C and MLO views. The third (difference) view was not used for further analysis in this study. Figure 2: Synthetic mammogram example. From left to right: A synthetic RGB mammogram after being normalized to 99%; C view extracted from the synthetic RGB image; MLO view extracted from the synthetic RGB image; |CC−MLO| C-MLO view extracted from the RGB image. Following the visual inspection, 94% of the synthetic images were judged as anatomically consistent, with only minor artifacts attributable to the blending of views. The remaining 6% exhibited major artifacts, primarily due to preprocessing limitations or inconsistencies in the original dataset. Table 1 summarizes the descriptive statistics for the IoU and DSC, comparing the binary masks of C and MLO views. The masks used for this analysis can be seen in Figure 3. Figure 4 presents the corresponding violin plots for each metric, illustrating the distribution across both synthetic and real samples. Table 1: Performance Statistics for Original and Synthetic Data Statistic Original Synthetic IoU DSC IoU DSC Mean 0.654 0.784 0.674 0.800 Mean difference (vs real) N/A N/A 0.020 0.016 Std 0.113 0.092 0.107 0.081 Min 0.000 0.000 0.261 0.414 Q1 (25%) 0.592 0.743 0.605 0.754 Median 0.668 0.801 0.682 0.811 Q3 (75%) 0.733 0.846 0.751 0.858 Max 0.932 0.964 0.908 0.925 IQR 0.141 0.102 0.146 0.103 Figure 3: Synthetic mammograms with mask. From left to right: C view; C mask; MLO view; MLO mask. Figure 4: IoU and DSC distributions for 2500 pairs of real and 500 pairs of synthetic mammograms’ masks. Additionally, Table 2 reports the results of two statistical tests—Kolmogorov-Smirnov (KS) and Earth Mover’s Distance (EMD)—used to compare the distributions of the real and synthetic datasets. These tests revealed statistically significant differences in IOU and DICE coefficient distributions (p<10−19p<10^-19), suggesting measurable variation between real and synthetic consistency. In contrast, the overlap metric also showed a statistically significant difference, though to a lesser extent (p<0.05p<0.05). Table 2: KS and EMD tests for IoU and DSC, comparing both real and synthetic mammograms (0.005<<p<<0.05). Statistic IoU DSC KS D 0.077∗0.077^* 0.077∗0.077^* EMD 0.020 0.015 4 DISCUSSION The results demonstrate that simultaneous dual-view synthesis is feasible using a diffusion-based framework and an RGB encoding strategy. Encoding C and MLO views together with their absolute difference allowed the model to learn cross-view anatomical relationships and produce paired images that were generally coherent in size, orientation, and breast shape. The consistency observed in the visual inspection suggests that joint generation is achievable even when starting from a pretrained natural-image DDPM, requiring only fine-tuning on mammograms. While prior studies on mammogram synthesis have shown strong performance, they have predominantly focused on generating single-view images or estimating the missing view from an existing one. GAN-based image-to-image models, for example, have demonstrated view-translation capabilities, whereas recent diffusion models have been used to generate high-quality full-field mammograms [13, 9, 5, 8]. However, these approaches do not address the challenge of ensuring anatomical consistency between C and MLO views of the same breast. In contrast, our approach directly encodes both projections within a unified representation, enabling the model to learn cross-view correspondences implicitly during training. Despite these encouraging results, the method still presents limitations. First, the absolute-difference channel may amplify small misalignments or intensity variations between C and MLO views, which could explain some of the artifacts seen in the synthetic samples. Additionally, the evaluation primarily focused on global shape consistency; future analyses should include lesion-level or density-specific assessments to better understand clinical realism. Finally, because this work is based on a single encoding strategy, additional encodings—or conditional mechanisms—may further improve fidelity and reduce artifacts. Overall, this preliminary study suggests that diffusion models hold promise for generating anatomically consistent dual-view mammograms, opening possibilities for data augmentation in dual-view CAD systems. Future work will explore alternative encodings, conditional generation, and external validation across multiple datasets. 5 CONCLUSIONS Dual-view mammogram generation is essential for advancing computer-aided diagnosis in breast cancer, as clinical interpretation often depends on the complementary information provided by the C and MLO views. In this study, we introduced a novel RGB-based encoding strategy, where each channel represents a distinct anatomical projection—C, MLO, and their absolute difference. This design enabled the model to learn a joint representation of both views and synthesize them simultaneously, promoting structural and size consistency. Although the generated images are not yet suitable for clinical deployment, the proposed method demonstrates promising results in terms of anatomical coherence and visual plausibility. Future work will focus on refining the training pipeline to reduce view-specific artifacts and improve image fidelity. We also plan to explore the effect of the third channel for the DDPM, as well as conditional generation using Latent Diffusion Models or Stable Diffusion frameworks, with conditioning variables such as BI-RADS category to enhance clinical relevance. Additionally, we aim to increase the diversity of the dataset by including more BI-RADS categories and to assess model generalizability through external validation across multiple datasets. Acknowledgements.This research was supported by the Secretaría de Ciencia, Humanidades, Tecnología e Innovación (Secihti), with cloud computing resources provided through Microsoft’s AI for Good Lab. References [1] M. Busaleh, M. Hussain, H. A. Aboalsamh, Fazal-e-Amin, and S. A. Al Sultan (2022) TwoViewDensityNet: two-view mammographic breast density classification based on deep convolutional neural network. Mathematics 10 (23). External Links: Link, ISSN 2227-7390, Document Cited by: §1. [2] J. Fitzjohn, C. Zhou, and J. G. Chase (2023-01) Critical assessment of mammography accuracy. IFAC-PapersOnLine 56, p. 5620–5625. External Links: Document, ISBN 9781713872344, ISSN 2405-8963, Link Cited by: §1. [3] J. A. Garza-Abdala, G. A. Fumagal-González, B. A. Bosques-Palomo, M. A. M. Molina, D. Avedano, S. Cardona-Huerta, and J. G. Tamez-Pena (2025) Ensemble of radiomics and convnext for breast cancer diagnosis. In 2025 IEEE 38th International Symposium on Computer-Based Medical Systems (CBMS), Vol. , p. 303–306. External Links: Document Cited by: §2.1. [4] J. Ho, A. Jain, and P. Abbeel (2020-06) Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems 2020-December. External Links: ISBN 2006.11239v2, ISSN 10495258, Link Cited by: §1, §2.3. [5] A. J. Joseph, P. Dwivedi, J. Joseph, S. Francis, P. P.N., J. P.B., A. V. Shamsu, and P. Sankaran (2024) Prior-guided generative adversarial network for mammogram synthesis. Biomedical Signal Processing and Control 87, p. 105456. External Links: ISSN 1746-8094, Document, Link Cited by: §4. [6] Z. Khan, M. Botlagunta, G. L. A. Kumari, P. Malviya, and M. Botlagunta (2024) Advancements in machine learning and deep learning for breast cancer detection: a systematic review. In Federated Learning, S. Ahmad, M. Alharbi, S. Jha, A. Ali, and R. Damaševičius (Eds.), External Links: Document, Link Cited by: §1. [7] B. Maistry and A. E. Ezugwu (2023-05) Breast cancer detection and diagnosis: a comparative study of state-of-the-arts deep learning architectures. External Links: Link Cited by: §1. [8] X. Meng and Y. Kabashima (2024) Diffusion model based posterior sampling for noisy linear inverse problems. External Links: 2211.12343, Link Cited by: §4. [9] R. Montoya-del-Angel, K. Sam-Millan, J. C. Vilanova, and R. Martí (2024) MAM-e: mammographic synthetic image generation with diffusion models. Sensors 24 (7). External Links: Link, ISSN 1424-8220, Document Cited by: §4. [10] R. Sutjiadi, S. Sendari, H. W. Herwanto, and Y. Kristian (2024) Generating high-quality synthetic mammogram images using denoising diffusion probabilistic models: a novel approach for augmenting deep learning datasets. In 2024 International Conference on Information Technology Systems and Innovation (ICITSI), Vol. , p. 386–392. External Links: Document Cited by: §1. [11] Z. Wang, J. Xian, K. Liu, X. Li, Q. Li, and X. Yang (2023) Dual-view correlation hybrid attention network for robust holistic mammogram classification. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI ’23. External Links: ISBN 978-1-956792-03-4, Link, Document Cited by: §1. [12] World Health Organization Breast cancer. Note: WHO, 26 March 2024 https://w.who.int/news-room/fact-sheets/detail/breast-cancer(Accessed: 29 July 2025) Cited by: §1. [13] A. Yamazaki and T. Ishida (2022-11) Two-view mammogram synthesis from single-view data using generative adversarial networks. Applied Sciences 2022, Vol. 12, Page 12206 12, p. 12206. External Links: Document, ISSN 2076-3417, Link Cited by: §1, §4.

Instant research discovery