Paper deep dive
ViPSAM: Visual Prompting Medical Image Segmentation Using Segment Anything Model
San Lee, Nalee Kim, Jeong Il Yu, Hee Chul Park, Boah Kim
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 7/17/2026, 5:22:01 AM
Summary
The paper introduces ViPSAM, a novel visual prompting framework for medical image segmentation built upon the Segment Anything Model (SAM). It addresses the challenge of segmenting low-contrast lesions in non-contrast CT (NCCT) images used for proton therapy planning by leveraging complementary information from contrast-enhanced MRI. ViPSAM incorporates a visual prompt encoder, a visual-guided cross-attention module, and Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning. Evaluated on a liver lesion segmentation dataset, ViPSAM demonstrates superior accuracy and robustness compared to baseline U-Net and SAM-based methods, highlighting the effectiveness of cross-modality visual prompting in clinical imaging workflows.
Entities (15)
Relation Signals (12)
ViPSAM → builtupon → Segment Anything Model (SAM)
confidence 97% · Built upon the Segment Anything Model (SAM), ViPSAM introduces a visual prompt encoder...
ViPSAM → targets → Liver Lesion Segmentation
confidence 97% · We evaluate the proposed method on liver lesion segmentation using NCCT acquired for proton therapy.
ViPSAM → uses → Visual Prompt Encoder
confidence 96% · we introduce a visual prompt encoder to extract soft-tissue contrast cues from contrast-enhanced images
ViPSAM → uses → Visual-Guided Cross-Attention Module
confidence 96% · visual-guided cross-attention module that conditions non-contrast image features on the extracted visual prompt representations
ViPSAM → achieves → Dice Similarity Coefficient
confidence 95% · segmentation performance was assessed using the Dice similarity coefficient (DSC)... ViPSAM MedSAM achieved the best liver segmentation performance with a Dice score of 0.937
ViPSAM → outperforms → U-Net
confidence 95% · Experimental results demonstrate that ViPSAM outperforms representative U-Net- and SAM-based methods
ViPSAM → uses → Low Rank Adaptation (LoRA)
confidence 95% · adapt the mask decoder by incorporating Low-Rank Adaptation (LoRA) for parameter-efficient fine-tuning
ViPSAM → →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:In proton therapy planning, respiratory-gated non-contrast CT (NCCT) is commonly used for lesion segmentation; however, accurate delineation remains challenging due to low lesion-to-background contrast. Although learning-based methods have shown strong performance, they often struggle with non-contrast image segmentation. Inspired by clinical practice, where contrast-enhanced MRI is referenced to delineate lesions on NCCT, we propose ViPSAM, a visual prompting framework that leverages complementary cross-modality information. Built upon the Segment Anything Model (SAM), ViPSAM introduces a visual prompt encoder to extract guidance features from contrast-enhanced images and a visual-guided cross-attention module to integrate non-contrast and contrast-enhanced features, thereby enhancing lesion-relevant representations in low-contrast regions. The mask decoder is further adapted in a parameter-efficient manner to utilize visual prompts effectively. We evaluate the proposed method on liver lesion segmentation using NCCT acquired for proton therapy. Experimental results demonstrate that ViPSAM outperforms representative U-Net- and SAM-based methods, indicating that cross-modality visual prompting enables more robust and accurate segmentation in non-contrast images.
Tags
Links
- Source: https://arxiv.org/abs/2607.14328v1
- Canonical: https://arxiv.org/abs/2607.14328v1
Trouble viewing inline? Open PDF directly →
Full Text
26,347 characters extracted from source content.
Expand or collapse full text
ViPSAM: Visual Prompting Medical Image Segmentation Using Segment Anything Model San Lee 1 , Nalee Kim 2 , Jeong Il Yu 2 , Hee Chul Park 2 , and Boah Kim 3⋆ 1 Department of Artificial Intelligence, Sungkyunkwan University, Republic of Korea 2 Department of Radiation Oncology, Samsung Medical Center, Sungkyunkwan University School of Medicine, Republic of Korea 3 Department of MetaBioHealth, Sungkyunkwan University, Republic of Korea Abstract. In proton therapy planning, respiratory-gated non-contrast CT (NCCT) is commonly used for lesion segmentation; however, accu- rate delineation remains challenging due to low lesion-to-background con- trast. Although learning-based methods have shown strong performance, they often struggle with non-contrast image segmentation. Inspired by clinical practice, where contrast-enhanced MRI is referenced to delin- eate lesions on NCCT, we propose ViPSAM, a visual prompting frame- work that leverages complementary cross-modality information. Built upon the Segment Anything Model (SAM), ViPSAM introduces a visual prompt encoder to extract guidance features from contrast-enhanced im- ages and a visual-guided cross-attention module to integrate non-contrast and contrast-enhanced features, thereby enhancing lesion-relevant repre- sentations in low-contrast regions. The mask decoder is further adapted in a parameter-efficient manner to utilize visual prompts effectively. We evaluate the proposed method on liver lesion segmentation using NCCT acquired for proton therapy. Experimental results demonstrate that ViP- SAM outperforms representative U-Net- and SAM-based methods, indi- cating that cross-modality visual prompting enables more robust and accurate segmentation in non-contrast images. Keywords: Medical image segmentation· Segment anything model· Visual prompt· Multi-modality 1 Introduction Accurate lesion delineation is important for proton therapy planning of liver tumors, as it determines target localization and proton beam delivery [5,14]. In clinical workflows, respiratory-gated non-contrast CT (NCCT) serves as the reference image for treatment planning [6]. However, precise segmentation on NCCT is challenging due to respiratory-induced anatomical variations and the inherently low contrast between lesions and surrounding liver regions [9,18,3,19]. To address this, clinicians often refer to contrast-enhanced MRI acquired in a ⋆ Corresponding author. Email: boah.kim at skku.edu arXiv:2607.14328v1 [eess.IV] 15 Jul 2026 2S. Lee et al. specific respiratory phase to better identify the lesion boundaries while contour- ing on NCCT. Nevertheless, the manual process is time-consuming and prone to inter-observer variability [16,13]. These challenges highlight the need for robust automated segmentation methods for low-contrast treatment planning images. With advances in deep learning, medical image segmentation has been exten- sively studied using U-Net-based architectures [15] and their extensions [10,4,2]. Although these models have achieved outstanding performance on various tasks, they typically operate in a single-modality setting and rely solely on intensity- based appearance cues. Accordingly, in non-contrast medical images, the low lesion-to-background contrast reduces feature discriminability, making it diffi- cult for models to segment lesion boundaries with high performance. Recently, foundation models trained on large-scale datasets have demon- strated strong transferability across downstream tasks, including image seg- mentation [11,21,17,7]. In particular, the Segment Anything Model (SAM) [11] presents a prompt-based segmentation framework and has shown remarkable generalization in medical imaging applications [12,20]. However, when applied to low-contrast images such as NCCT, SAM-based models still struggle to de- lineate small lesions with indistinct boundaries. Moreover, adapting foundation models to specific medical domains often requires substantial computational cost. To address these challenges, we propose a visual prompting medical image segmentation model based on SAM, termed ViPSAM, which leverages comple- mentary multi-modal information within a unified framework. Inspired by the clinical practice where radiation oncologists reference contrast-enhanced MRI to delineate lesions on NCCT, our model incorporates cross-modality visual guid- ance to enhance segmentation in low-contrast clinical scenarios. Specifically, we introduce a visual prompt encoder to extract soft-tissue con- trast cues from contrast-enhanced images, which serve as visual prompts and provide complementary information to non-contrast medical images. These cues are integrated with non-contrast image representations through a visual-guided cross-attention module, enabling the model to better capture lesion-relevant information. Furthermore, we adapt the SAM mask decoder by incorporating Low-Rank Adaptation (LoRA) [8] for parameter-efficient fine-tuning. Through this design, segmentation is performed on non-contrast images, while contrast- enhanced images are used solely as visual prompts. We evaluate ViPSAM for liver lesion segmentation using a proton therapy dataset collected at a tertiary medical center, consisting of NCCT and cor- responding contrast-enhanced MRI. Segmentation is performed on NCCT us- ing MRI-derived visual prompts. Experimental results demonstrate that visual prompting significantly improves segmentation performance, surpassing repre- sentative U-Net- and SAM-based methods. Our contributions are as follows: – We propose ViPSAM, a SAM-based visual prompting framework leveraging cross-modal cues for lesion segmentation in non-contrast medical images. – We design a visual prompt encoder and a visual-guided cross-attention mod- ule that conditions non-contrast features on contrast-enhanced visual prompts, together with LoRA-based parameter-efficient adaptation within SAM. ViPSAM: Visual Prompting Medical Image Segmentation Using SAM3 Fig. 1. (a) Overview of ViPSAM. Lesion segmentation ˆy on non-contrast images I nc is predicted by leveraging complementary information from contrast-enhanced images I ce via a visual prompt encoder, a visual-guided cross-attention (VGCA) module, and LoRA-based adaptation in the mask decoder. (b) Architecture of the VGCA module. – Experimental results on a liver lesion segmentation dataset for proton ther- apy with patient-matched NCCT and contrast-enhanced MRI demonstrate superior performance over representative U-Net- and SAM-based methods. 2 Proposed Method An overview of ViPSAM, our SAM-based framework for visual-prompting image segmentation, is shown in Fig. 1. ViPSAM extends the Segment Anything Model (SAM) by incorporating (i) a visual prompt encoder that extracts soft-tissue contrast cues from cross-modality contrast-enhanced images, (i) a visual-guided cross-attention module that conditions non-contrast image features on the ex- tracted visual prompt representations, and (i) Low-Rank Adaptation (LoRA) within the decoder for parameter-efficient fine-tuning. We first provide a brief overview of SAM, followed by detailed descriptions of each component. Segment Anything Model (SAM) SAM [11] consists of an image encoder E image , a prompt encoder E prompt , and a mask decoder D mask . The image en- coder, implemented as a Vision Transformer (ViT), takes an input image I and extracts image embedding F image = E image (I). The prompt encoder embeds a prompt P (e.g., points, boxes, or masks) into prompt tokens E prompt (P). The mask decoder then takes the image embedding F image and the prompt tokens E prompt (P) to predict a segmentation mask ˆy = D mask (F image ,E prompt (P)) via Two-Way Transformer blocks that perform bidirectional cross-attention between the prompt tokens and image embedding. In our framework, we adopt SAM with a ViT-B image encoder and keep its pretrained parameters frozen during train- ing, using box prompts as input. 4S. Lee et al. 2.1 Visual Prompt Encoder To mitigate low lesion-to-background contrast in non-contrast images, we intro- duce a visual prompt encoder E visual prompt , where visual prompt refers to image-based guidance from contrast-enhanced images that provide clearer lesion contrast, in contrast to sparse prompts (e.g., boxes). Notably, E visual prompt uses the same ViT-B architecture and SAM-pretrained weights as the image encoder E image , with all weights frozen during training. Given a non-contrast image I nc and a contrast-enhanced image I ce , each image is replicated to three channels to match the SAM input format, yielding I nc ,I ce ∈R 3×H×W , where H and W represent the height and width of the input image, respectively. When the non-contrast image is fed into E image to extract: F image = E image (I nc )∈R c×h×w ,(1) where c and h× w denote the channel dimension and spatial resolution of the feature map, respectively, the contrast-enhanced image is encoded by E visual prompt to obtain visual prompt features: F visual = E visual prompt (I ce )∈R c×h×w .(2) Since both encoders have an identical architecture, the resulting feature maps can interact directly in the subsequent visual-guided cross-attention module. 2.2 Visual-Guided Cross-Attention Module To obtain visual-prompt-guided representations with the non-contrast image serving as the spatial reference, we design a visual-guided cross-attention (VGCA) module (Fig. 1(b)) that conditions non-contrast image features F image on visual prompt features F visual : X V GCA = VGCA(F image ,F visual ).(3) Specifically, in the VGCA module, F image and F visual are first flattened and normalized by layer normalization (LN) [1]: X image = LN(Flat(F image )), X visual = LN(Flat(F visual )),(4) where Flat(·) reshapesR c×h×w toR n×c with n = h·w. Then, multi-head cross- attention (MHA) is computed as: A = MHA(Q = X image ,K = X visual ,V = X visual ),(5) where Q, K, and V denote the query, key, and value, respectively. Since segmen- tation is performed on non-contrast images, X image is used as the query, while X visual serves as the key and value to provide complementary lesion information from contrast-enhanced images. ViPSAM: Visual Prompting Medical Image Segmentation Using SAM5 Fig. 2. LoRA placement in the SAM mask decoder. (a) Two-Way Transformer block in the mask decoder. (b,c) LoRA is applied to selected projections in cross-attention layers: token-to-image (K, V, O) and image-to-token (Q, O). To regulate the influence of visual prompts, a learnable gating scalar γ is introduced through a gated residual update: X γ = LN(X image + γA),(6) ̃ X = LN(X γ + MLP(X γ )),(7) where MLP denotes a feed-forward network with a Linear–GeLU–Linear struc- ture, and ̃ X is the cross-attended representation. The visual-prompt-conditioned image embedding X V GCA , used in the mask decoder, is then computed as: X V GCA = λX image + (1− λ) ̃ X,(8) where λ is a learnable weight scalar. Unlike standard cross-attention, our for- mulation with learnable gating and weighting preserves the non-contrast image as the primary spatial reference while enabling cross-modal visual guidance to modulate feature representations. 2.3 LoRA in Mask Decoder In our framework, the mask decoder D mask consists of stacked Two-Way Trans- former blocks that alternate token-to-image and image-to-token cross-attention (Fig. 2(a)). This takes sparse prompt tokens generated by the sparse prompt en- coder E sparse prompt from a box prompt P which provides localization cues, together with image embeddings produced by the VGCA module, and predicts the seg- mentation mask ˆy by: ˆy = D mask (X V GCA ,E sparse prompt (P)).(9) 6S. Lee et al. Here, to effectively adapt the mask decoder to visual-prompt-conditioned im- age embeddings X V GCA in a parameter-efficient manner, we incorporate Low- Rank Adaptation (LoRA) [8] into the decoder by updating frozen linear projec- tions with learnable low-rank residuals. Details are described below. Let X ∈R n×c in denote an image embedding and W ∈R c out ×c in be a frozen linear projection producing ˆ X = WX. In the LoRA layer, the adapted weight is given by: ̃ W = W + BA,(10) where A ∈R r×c in and B ∈R c out ×r are learnable linear layer with rank r ≪ min(c in ,c out ). During training, W remains fixed and only (A,B) are updated, enabling parameter-efficient adaptation with minimal additional parameters. Following this strategy, we insert LoRA into selected cross-attention pro- jections of the mask decoder according to the role of image embeddings in each pathway. Specifically, in token-to-image (t2i) attention (Fig. 2(b)), sparse prompt tokens serve as queries (Q) to retrieve spatially relevant image cues for target lo- calization, while image embeddings act as keys (K) and values (V ). Since image embeddings are projected into K and V , LoRA is applied to the key, value, and output projection matrices, i.e., ̃ W t2i K , ̃ W t2i V , and ̃ W t2i O . In contrast, in image-to- token (i2t) attention (Fig. 2(c)), image embeddings serve as queries to attend back to sparse prompt tokens for feature refinement, while sparse prompt tokens act as keys and values. Accordingly, LoRA is applied to the query and output projection matrices, i.e., ̃ W i2t Q and ̃ W i2t O . Together, the t2i and i2t attention layers enable bidirectional interaction be- tween sparse prompts and dense image features, while selective LoRA placement facilitates efficient adaptation of the mask decoder to visual-prompt-conditioned representations without full fine-tuning, resulting in accurate lesion segmenta- tion on non-contrast images. 3 Experiments Dataset and Metric To evaluate the proposed cross-modality visual prompt- ing framework, we used a liver lesion dataset for proton therapy planning col- lected at Samsung Medical Center. The dataset comprises 73 cases, each includ- ing a mid-respiratory phase NCCT (T = 50%), a T1-weighted fat-suppressed contrast-enhanced MRI at the same respiratory phase, and expert-annotated liver and lesion masks defined on the NCCT. MRI scans were rigidly registered to the NCCT, and all scans were resampled to a voxel spacing of 0.6597×0.6597×5 m. NCCT scans were clipped to [-100, 200] HU. For both CT and MRI, each 2D axial slice was resized to 1024×1024 and normalized to [0, 1] using min-max scaling. The dataset was split into 60, 6, and 7 cases for training, validation, and testing, respectively. In total, 2,251, 213, and 245 2D axial slices containing liver lesions were used for each split. For quantitative evaluation, segmentation per- formance was assessed using the Dice similarity coefficient (DSC), Intersection over Union (IoU), and the 95th percentile Hausdorff distance (HD95). ViPSAM: Visual Prompting Medical Image Segmentation Using SAM7 Fig. 3. Qualitative comparison results of liver lesion segmentation on non-contrast CT. Red denotes the liver and green denotes the liver lesion. Table 1. Quantitative comparison results of the average Dice, IoU, HD95 and the number of trainable parameters. Bold indicates the best performance for each metric. Methods LiverLesion Params (M) ↓ Dice ↑IoU ↑HD95(m) ↓ Dice ↑IoU ↑HD95(m) ↓ U-Net0.875±0.162 0.805±0.187 27.62±31.13 0.236±0.298 0.174±0.240 60.42±31.35 4.32 nnU-Net0.856±0.238 0.799±0.244 28.72±41.09 0.267±0.339 0.209±0.281 43.23±28.12 59.18 TransUNet0.861±0.198 0.793±0.219 22.29±26.83 0.169±0.189 0.106±0.132 77.92±46.17 105.28 Swin-Unet0.880±0.155 0.809±0.178 24.38±26.85 0.162±0.189 0.101±0.128 56.25±25.63 27.17 SAM0.890±0.111 0.815±0.138 18.80±13.27 0.655±0.208 0.520±0.215 13.11±10.02 93.74 MedSAM0.892±0.105 0.817±0.136 13.59±12.29 0.709±0.180 0.577±0.206 8.39±4.76 93.74 SAMed0.901±0.128 0.838±0.155 18.88±27.22 0.341±0.328 0.259±0.270 48.59±28.12 3.93 ViPSAM SAM 0.925±0.088 0.869±0.114 14.58±17.68 0.852±0.114 0.757±0.150 5.15±3.70 0.83 ViPSAM MedSAM 0.937±0.070 0.887±0.094 8.84±9.49 0.810±0.151 0.704±0.188 5.18±2.18 0.83 Implementation Details We implemented the proposed ViPSAM using either SAM [11] (ViPSAM SAM ) or MedSAM [12] (ViPSAM MedSAM ) as the backbone, each initialized with pretrained weights. For the VGCA module, the feature channel dimension was set to C = 256 and the number of attention heads was set to 8. We initialize the gating scalar γ to 0.3, the weight λ to 1.0, and the LoRA rank r to 8. Box prompts were generated from manual masks with random perturbations (≤20 px) during training and fixed during inference. The model was trained for 12 epochs using an equally weighted sum of Dice and binary cross-entropy losses and optimized with AdamW with a learning rate of 1 × 10 −4 and a weight decay of 1 × 10 −4 . The checkpoint from the final epoch was used for testing. All experiments were conducted in PyTorch 2.5.1 on an NVIDIA RTX A6000 GPU. Memory usage was approximately 14.3GB during training and 4.8GB during inference. The source code is publicly available at https://github.com/torchViPSAM/ViPSAM. 8S. Lee et al. Table 2. Ablation study of ViPSAM on prompt design and trainable components. Config. PromptLayerLiverLesion Visual Sparse CA LoRA Dice ↑IoU ↑ HD95(m) ↓ Dice ↑IoU ↑ HD95(m) ↓ (a)✓ 0.909±0.091 0.842±0.119 12.04±11.02 0.801±0.173 0.695±0.187 5.84±3.14 (b)✓ 0.809±0.236 0.728±0.247 51.85±44.99 0.155±0.176 0.095±0.116 105.91±43.41 (c)✓0.932±0.081 0.881±0.108 9.62±10.68 0.808±0.136 0.669±0.181 5.04±1.87 Ours✓ 0.937±0.070 0.887±0.094 8.84±9.49 0.810±0.151 0.704±0.188 5.18±2.18 Experimental Results We compared ViPSAM with representative segmenta- tion models, including U-Net-based methods (U-Net [15], nnU-Net [10], Tran- sUNet [4], and Swin-Unet [2]) and SAM-based methods using box prompts (SAM [11], MedSAM [12], and SAMed [20]). Baseline models were trained using only NCCT blue with the default settings provided in the original papers, whereas ViPSAM additionally leveraged contrast-enhanced MRI as visual prompts. All methods were evaluated on NCCT for liver and liver lesion segmentation. Fig. 3 visualizes qualitative comparisons on NCCT. ViPSAM produces more accurate segmentation masks than the existing SAM-based models, even for small lesions with indistinct boundaries. As reported in Table 1, the proposed ViPSAM consistently outperformed both U-Net- and SAM-based methods across all evaluation metrics. Specifically, ViPSAM MedSAM achieved the best liver seg- mentation performance with a Dice score of 0.937 and an HD95 of 8.84 m, while ViPSAM SAM achieved the best lesion segmentation performance with a Dice score of 0.852 and an HD95 of 5.15 m. In contrast, U-Net-based models showed poor lesion segmentation performance under low-contrast NCCT con- ditions (Dice < 0.267), and SAM-based methods exhibited imprecise boundary delineation for lesions despite box-prompt guidance. Also, our method required only 0.83M trainable parameters, highlighting superior parameter efficiency. Moreover, we assessed statistical significance using paired t-tests and Wilcoxon signed-rank tests. Both ViPSAM SAM and ViPSAM MedSAM significantly out- performed their respective baselines (SAM and MedSAM) across all metrics (p < 0.05) under both tests. These results support the effectiveness of our model leveraging MRI-derived visual prompting for liver lesion segmentation on NCCT. Ablation Study To evaluate each component in ViPSAM, we conducted an ablation study with three configurations: (a) ViPSAM without visual prompting, (b) ViPSAM without sparse box prompts, and (c) ViPSAM without LoRA- based adaptation. All configurations were trained under the same settings for fair comparison. Table 2 shows quantitative results for liver and lesion segmentation across the ablated models. Without visual prompting (a), performance decreased to a lesion Dice of 0.801 and an IoU of 0.695, compared to the full ViPSAM (Dice=0.810, IoU=0.704), highlighting the importance of incorporating cross-modal contrast information. Moreover, removing sparse box prompts (b) led to further per- formance degradation, suggesting their role in stabilizing lesion localization. In addition, without LoRA-based adaptation (c), the model showed reduced per- ViPSAM: Visual Prompting Medical Image Segmentation Using SAM9 formance with a Dice of 0.808 and an IoU of 0.669. These results indicate that visual prompting with contrast-enhanced images, VGCA-based cross-modal in- teraction, and LoRA-based decoder adaptation contribute synergistically to im- proved liver lesion segmentation on NCCT. 4 Conclusion In this work, we propose ViPSAM, a SAM-based visual prompting framework for non-contrast medical image segmentation. By leveraging complementary in- formation from contrast-enhanced images via a visual prompt encoder, visual- guided cross-attention, and LoRA-based parameter-efficient decoder adaptation, ViPSAM enables precise segmentation with minimal additional learnable pa- rameters. Experimental results on a liver lesion proton therapy planning dataset demonstrate the effectiveness of the proposed framework in alleviating bound- ary ambiguity in non-contrast imaging, highlighting its potential to enhance segmentation reliability in clinical workflows. Acknowledgments. This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (RS-2026-25481239), AI Graduate School Support Program (Sungkyunkwan University) (RS-2019-I190421), and SKKU Academic Research Support Program, Sungkyunkwan University, 2025. Disclosure of Interests. The authors have no competing interests to declare that are relevant to the content of this article. References 1. Ba, J.L., Kiros, J.R., Hinton, G.E.: Layer normalization. arXiv preprint arXiv:1607.06450 (2016) 2. Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swin- unet: Unet-like pure transformer for medical image segmentation. In: European conference on computer vision. p. 205–218. Springer (2022) 3. Chang, J.Y., Zhang, X., Knopf, A., Li, H., Mori, S., Dong, L., Lu, H.M., Liu, W., Badiyan, S.N., Both, S., et al.: Consensus guidelines for implementing pencil-beam scanning proton therapy for thoracic malignancies on behalf of the ptcog thoracic and lymphoma subcommittee. International Journal of Radiation Oncology* Biol- ogy* Physics 99(1), 41–50 (2017) 4. Chen, J., Lu, Y., Yu, Q., Luo, X., Adeli, E., Wang, Y., Lu, L., Yuille, A.L., Zhou, Y.: Transunet: Transformers make strong encoders for medical image segmentation. arXiv preprint arXiv:2102.04306 (2021) 5. Cheng, J.Y., Liu, C.M., Wang, Y.M., Hsu, H.C., Huang, E.Y., Huang, T.T., Lee, C.H., Hung, S.P., Huang, B.S.: Proton versus photon radiotherapy for primary hep- atocellular carcinoma: a propensity-matched analysis. Radiation Oncology 15(1), 159 (2020) 6. Cheung, A.L.Y., Zhang, L., Liu, C., Li, T., Cheung, A.H.Y., Leung, C., Leung, A.K.C., Lam, S.K., Lee, V.H.F., Cai, J.: Evaluation of multisource adaptive mri fusion for gross tumor volume delineation of hepatocellular carcinoma. Frontiers in oncology 12, 816678 (2022) 10S. Lee et al. 7. Cox, J., Liu, P., Stolte, S.E., Yang, Y., Liu, K., See, K.B., Ju, H., Fang, R.: Brain- segfounder: Towards 3d foundation models for neuroimage segmentation. Medical Image Analysis 97, 103301 (2024) 8. Hu, E.J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., Chen, W., et al.: Lora: Low-rank adaptation of large language models. ICLR 1(2), 3 (2022) 9. Iannaccone, R., Laghi, A., Catalano, C., Rossi, P., Mangiapane, F., Murakami, T., Hori, M., Piacentini, F., Nofroni, I., Passariello, R.: Hepatocellular carcinoma: role of unenhanced and delayed phase multi–detector row helical ct in patients with cirrhosis. Radiology 234(2), 460–467 (2005) 10. Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods 18(2), 203–211 (2021) 11. Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF international conference on computer vision. p. 4015–4026 (2023) 12. Ma, J., He, Y., Li, F., Han, L., You, C., Wang, B.: Segment anything in medical images. Nature Communications 15(1), 654 (2024) 13. Marshall, C., Thirion, P., Mihai, A., Armstrong, J.G., Cournane, S., Hickey, D., McClean, B., Quinn, J.: Interobserver variability of gross tumor volume delineation for colorectal liver metastases using computed tomography and magnetic resonance imaging. Advances in Radiation Oncology 8(1), 101020 (2023) 14. Newhauser, W.D., Zhang, R.: The physics of proton therapy. Physics in Medicine & Biology 60(8), R155 (2015) 15. Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. p. 234–241. Springer (2015) 16. Sherer, M.V., Lin, D., Elguindi, S., Duke, S., Tan, L.T., Cacicedo, J., Dahele, M., Gillespie, E.F.: Metrics to evaluate the performance of auto-segmentation for radiation treatment planning: A critical review. Radiotherapy and Oncology 160, 185–191 (2021) 17. Shi, P., Qiu, J., Abaxi, S.M.D., Wei, H., Lo, F.P.W., Yuan, W.: Generalist vision foundation models for medical imaging: A case study of segment anything model on zero-shot medical segmentation. Diagnostics 13(11), 1947 (2023) 18. Tsai, Y.L., Wu, C.J., Shaw, S., Yu, P.C., Nien, H.H., Lui, L.T.: Quantitative anal- ysis of respiration-induced motion of each liver segment with helical computed tomography and 4-dimensional computed tomography. Radiation Oncology 13(1), 59 (2018) 19. Velec, M., Moseley, J.L., Eccles, C.L., Craig, T., Sharpe, M.B., Dawson, L.A., Brock, K.K.: Effect of breathing motion on radiotherapy dose accumulation in the abdomen using deformable registration. International Journal of Radiation Oncology* Biology* Physics 80(1), 265–272 (2011) 20. Zhang, K., Liu, D.: Customized segment anything model for medical image seg- mentation. arXiv preprint arXiv:2304.13785 (2023) 21. Zhang, Y., Zhou, T., Wang, S., Liang, P., Zhang, Y., Chen, D.Z.: Input augmenta- tion with sam: Boosting medical image segmentation with segmentation foundation model. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. p. 129–139. Springer (2023)