Pith. sign in

REVIEW 4 major objections 5 minor 43 references

SARFA: Segment Anything with Radiomic Feature Alignment

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SARFA claims that training a medical segmenter to rank its candidate masks by radiomic distance to ground truth—rather than by pixel overlap alone—yields more clinically consistent segmentations, with best-in-class results on lung CT and br

desk verdict Promising idea—FRD as a DPO ranking signal for SAM candidates—but the reported gains need matched checkpoint selection, error bars, and a beta fix. read the letter →

arxiv 2607.13323 v1 pith:AKGF5WWL submitted 2026-07-14 cs.CV

Tyler Ward , Abdullah Imran This is my paper
classification cs.CV
keywords ambiguousmedicalsegmentationradiomicsFréchetRadiomicDistanceDirectPreferenceOptimizationSegmentAnythingModelprobabilisticpromptinglungCTbrainMRI
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that radiomic feature similarity can serve as a training objective for ambiguous medical image segmentation, not just as an evaluation metric. It introduces SARFA, which generates multiple candidate masks from SAM's multimask decoding, ranks them by Fréchet Radiomic Distance to the ground truth's masked-region features, and applies a DPO-style preference loss so the model learns to favor radiomically close masks. On lung CT and brain MRI benchmarks, the authors report that this combination improves all measured metrics over ten prior methods. If the claim is right, clinicians could train segmenters to respect texture and shape characteristics that matter for diagnosis, not just boundary pixels.

What carries the argument

The central object is the Fréchet Radiomic Distance (FRD), a 2-Wasserstein distance between Gaussian fits to hand-crafted radiomic feature distributions extracted from masked image regions. SARFA uses FRD in two ways: per-sample, it ranks the K candidate masks produced by SAM's multimask decoding—the closest to ground truth becomes the preferred mask, the farthest becomes the rejected mask—and at epoch level, the lowest-FRD checkpoint is saved. The DPO-style loss pushes the model's IoU-scoring head to prefer the radiomically close candidate, while a learnable weighted sum of candidate masks produces the final segmentation, preserving SAM's ambiguity while allowing task-specific scale and str

What would settle it

Train the exact SARFA pipeline with the DPO loss detached from the IoU-scoring head (or with the scoring head frozen) while keeping the supervised losses and mask-weighting update. If the reported GED/FRD values on LIDC (0.194 / 2.818) do not degrade, then the radiomic preference signal is not responsible for the improvement.

Watch

Extended reading notes

Core claim

The central claim is that radiomic feature alignment—quantified by Fréchet Radiomic Distance (FRD) between a predicted mask's radiomics and the ground-truth mask's radiomics—can be an effective training signal for ambiguous medical segmentation. SARFA generates K candidate masks from SAM's multimask decoder, extracts hand-crafted radiomic features from the masked regions, z-scores them using ground-truth statistics, and labels the candidate closest to the ground-truth radiomics as preferred and the farthest as rejected. A Direct Preference Optimization loss then trains SAM's IoU-scoring head to assign higher probability to the radiomically close mask, while a learned weighted sum of the K ca

Load-bearing premise

The results stand or fall on the assumption that the radiomic ranking actually steers the final prediction: the DPO loss updates only SAM's IoU-scoring head, while the output is a weighted average of candidate masks using separate learnable weights, and no experiment isolates whether updating that scoring head changes the final mask.

Editorial extensions

If this is right

  • If radiomic alignment genuinely drives performance, hand-crafted clinical texture and shape descriptors can be used as a training signal wherever pixel boundaries are ambiguous, not just as an evaluation metric.
  • Preference optimization over candidate masks transfers from ranking text to ranking dense segmentations, so other multi-output foundation models could be steered by non-overlap rewards.
  • The reported correlation between FRD and GED, Dice, and Hungarian-matched IoU means radiomic distance can act as a practical selection criterion during training and checkpointing.
  • Demonstrated gains on both lung CT and brain MRI suggest the mechanism is not tied to a single modality or anatomy.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves open whether the DPO loss is the causal driver: it updates SAM's IoU-scoring head while the final prediction is a weighted sum of candidate masks using separate learnable weights. A direct ablation that freezes the scoring head, or that trains the mask weights with the DPO loss, would settle this.
  • Because FRD is computed from masked-region features rather than full images, the same ranking could be applied to unlabeled candidates by comparing against a population-level radiomic profile of the target lesion type, potentially enabling semi-supervised training.
  • The paper uses only the best and worst candidate per image for preference construction; using the full ordering of all K candidates as multiple pairwise preferences is a natural variant that could provide denser gradients.
  • The authors list computational overhead, reliance on one radiomics implementation, and evaluation on two datasets as limitations; scaling to 3D volumes with an approximate FRD would test whether the objective remains practical.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes SARFA, a framework that adapts SAM to ambiguous medical image segmentation by generating multiple candidate masks via probabilistic prompting, weighting them with learnable scalar weights, and training with a combined supervised loss plus a DPO-style ranking loss whose preference pairs are selected by Fréchet Radiomic Distance (FRD) between candidate-mask radiomic features and ground-truth radiomic features. The model is selected on validation FRD. The authors evaluate on LIDC-IDRI (CT) and BraTS2017 (MRI), reporting that SARFA outperforms ten SOTA baselines on LIDC and both P2SAM variants on BraTS, and they include ablations over loss components, ranking strategies, and hyperparameters.

Significance. If the claims hold, the paper makes a useful contribution by introducing radiomic feature alignment as a training signal for a foundation segmentation model, going beyond pixel-level overlap losses. The work is timely, the code is promised, and the ablations in Tables 3 to 5 provide a useful starting point. However, the central empirical claim is currently not established to the standard normally required: the checkpoint-selection protocol is asymmetric between SARFA and the baselines in Table 1, the same FRD metric is used as training objective, model-selection criterion, and evaluation metric, the DPO gradient path to the final mask output is underspecified, and no error bars or significance tests are reported.

major comments (4)
  1. [§3.2, §4.3, Tables 1-2] The LIDC comparison is threatened by asymmetric checkpoint selection. Section 3.2 states that the checkpoint with the lowest epoch-level FRD is saved and used for evaluation. Table 2 explicitly distinguishes P2SAM(†) (lowest validation loss) from P2SAM(‡) (lowest FRD), but Table 1 provides no such notation for any baseline. Since Fig. 5 shows that FRD correlates strongly with GED, HM-IoU, and Dmax (|r|>0.9), selecting SARFA's checkpoint on validation FRD also indirectly optimizes the other reported metrics. The reported GED improvement over P2SAM (0.353 vs 0.206) may therefore reflect model-selection protocol rather than the proposed radiomic ranking. Please report all baselines under a matched checkpoint-selection rule, or at least provide both variants, and include error bars and significance tests over repeated runs.
  2. [§4.2 vs. Table 3 and tuning text] There is a direct contradiction in the DPO temperature. Section 4.2 says 'we used a β of 0.01', while the hyperparameter-tuning paragraph states that β=0.1 achieves the best balanced performance and 'is the configuration used for our experiments reported in Tables 1 and 2'. Table 3 also lists configurations with β=0.1 and β=0.01. This must be corrected and the actual configuration used in Tables 1-2 stated unambiguously.
  3. [§3.3, Eqs. (15)-(17), §4.3 Table 5] The mechanism linking the DPO loss to the final segmentation output is underspecified. Eq. (15)-(16) place DPO on the IoU logits s_theta of SAM's mask decoder, and the final prediction is the weighted sum of candidate masks using separate learnable weights w_k (Eqs. (5)-(6)). The paper never states whether gradients from L_DPO flow into the shared mask-decoder trunk (and hence alter the candidate masks) or only into the IoU prediction head. The ablations in Table 5 remove the entire DPO component, which does not isolate the logit pathway. If the DPO loss cannot change the candidate masks or the final weighted output, the reported gains cannot be attributed to the radiomic ranking; if it does, the text should describe the architectural gradient path explicitly.
  4. [§4.3, Tables 3-5 and Fig. 5] The ablation and correlation evidence is not yet sufficient to support the claim that FRD-based DPO ranking is a valid training objective. Tables 3-5 report single numbers without error bars or number of seeds, and it is unclear whether the ablations are evaluated on a validation split or the held-out test split. Given that checkpoint selection is FRD-based on the validation set, a single-seed ablation can reflect selection noise. Additionally, Fig. 5 shows that FRD correlates with overlap metrics across candidate masks produced during training; this is evidence that FRD is a reasonable proxy, but it does not establish that optimizing FRD as a loss improves generalization. Please provide repeated-seed statistics and state the evaluation split used for ablations.
minor comments (5)
  1. [Abstract / Section 1] There are several typos and broken sentences, e.g., 'SARF A' in the abstract, and in Section 4.3: 'As evidenced by the first of P2SAM's predicted masks, SARFA can accurately which completely fails to properly segment the lung lesion' — the sentence is incomplete.
  2. [Table 1] Many entries in Table 1 are dashes, so it is difficult to compare methods across all metrics. Please provide complete numbers for all baselines or explain why a metric is not applicable.
  3. [Algorithm 1] Algorithm 1 line 9 hardcodes M={m_hat_1,...,m_hat_8}, while Section 4.2 states K=16 candidate masks for the main experiments and K=8 for the tuning experiments. Please make the algorithm consistent with the main configuration.
  4. [Fig. 2 caption] The caption says 'the selection of this value' without naming the value; it should state that 10 epochs is selected.
  5. [Eq. (4)] The notation E_{tilde Y ~ P_theta(·|x)}[tilde Y] ≈ (1/K) Σ_k y_hat^(k) is a heuristic approximation; please clarify whether the candidate masks are treated as samples from a distribution and how the weights w_k relate to this approximation.

Circularity Check

1 steps flagged · score 5.0 of 10

FRD-based checkpoint selection makes the LIDC FRD comparison partly a selection artifact; the BraTS matched comparison and independent GED/HM-IoU results keep the central claim from being fully circular.

  1. fitted input called prediction [Section 3.2 (epoch-level FRD checkpoint selection); Section 4.3 Table 1; Figure 5]
    "In our approach, this epoch-level FRD calculation is used to determine which model checkpoint to save, with the checkpoint with the lowest FRD being the one saved and used for evaluation. ... Lower FRD values are associated with lower GED and higher Dice and HM-IoU scores, indicating that radiomic similarity remains strongly aligned with conventional overlap- and distribution-based segmentation metrics."

    The deployed SARFA checkpoint is chosen by minimizing validation FRD, and Table 1 then reports test FRD as one of the headline metrics. The paper itself claims FRD is highly correlated with GED, HM-IoU, and Dmax (|r|>0.9 in Fig. 5), so the same selection rule also favors the other reported metrics. Table 2 labels P2SAM(†)/(‡) by selection rule, but Table 1 gives no such labeling for P2SAM or the other baselines; hence the LIDC comparison is not matched by construction. The LIDC FRD result, and through the claimed correlation the neighboring GED/HM-IoU/Dmax margins, are partially an artifact of the checkpoint-selection criterion rather than an independent test of the radiomic-DPO objective.

full rationale

The paper's central derivation is mostly self-contained: FRD is defined from PyRadiomics features (Eqs. 7-13), preference pairs are constructed from per-mask radiomic distance (Eqs. 10-11), and the DPO loss is explicit (Eqs. 15-16). The self-citations [32]-[34] are related-work mentions and are not load-bearing; no uniqueness theorem is imported from the authors. The BraTS2017 comparison (Table 2) is more informative because it includes a P2SAM variant selected by lowest FRD (P2SAM(‡)); against that matched baseline SARFA still improves GED (2.644 vs 3.979) and other metrics, providing independent content. The main circular component is the LIDC FRD checkpoint selection described in Section 3.2: the reported FRD in Table 1 is the same criterion used to choose the checkpoint, and Fig. 5 is then used to extend that selection advantage to GED/HM-IoU/Dmax. This is a partial reduction of the headline comparison to the selection rule, but it is not a complete equivalence because the test set is held out and the non-FRD metrics are distinct quantities. A separate, non-circular weakness is that the DPO loss is placed on SAM's IoU logits (Eqs. 15-16) while the final output is the weighted sum of masks (Eq. 6); the paper never demonstrates that updating s_theta changes the final prediction, which weakens the causal attribution of the ablation results but is not a circularity. Overall, the central claim retains enough external comparison to avoid a 6+, but the asymmetric FRD-based checkpoint selection justifies a moderate score of 5.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The central claim depends on several domain assumptions: that radiomic features on 2D slices capture clinically meaningful structure, that SAM's multimask outputs are a valid approximation to a segmentation distribution, and that a DPO loss on logits can improve the actual masks. The DPO-logit assumption is the weakest: the paper does not trace a gradient or control path from s_theta to the final output (Eq. 6).

free parameters (6)
  • K (number of candidate masks) = 8 or 16
    Set by hand; K=16 used for main results (Tables 1-2), K=8 for ablation (Table 3).
  • beta (DPO temperature) = 0.01 (Sec. 4.2) or 0.1 (Sec. 4.3)
    Empirically tuned; conflicting values reported for the same headline experiments.
  • lambda (DPO loss weight) = 0.05
    Chosen by hyperparameter sweep (Table 3).
  • T (DPO interval in steps) = 10
    Chosen from sweep; results are highly sensitive (GED 0.194 at T=10, 6.502 at T=15).
  • training epochs = 10
    Selected from loss-curve convergence in Fig. 2; no metric-based early stopping.
  • learning rate = 0.001
    Standard choice, not justified.
assumptions (4)
  • domain assumption PyRadiomics features extracted from 2D masked regions provide clinically meaningful descriptors for segmentation quality.
    The whole FRD ranking relies on this; Fig. 5 supports correlation with overlap metrics over 80 training candidates only.
  • domain assumption SAM's built-in multimask outputs can be treated as samples from an implicit segmentation distribution (Eq. 3-4) without explicit probabilistic prompting.
    This replaces P2SAM's Gaussian prompt sampling and is asserted, not derived.
  • ad hoc to paper The DPO loss on IoU logits improves the quality of generated candidate masks.
    Not established; final prediction uses w_k, not s_theta; ablation does not isolate this path.
  • standard math The 2-Wasserstein distance between Gaussian approximations (FRD, Eq. 13) is an appropriate metric for radiomic feature-set comparison.
    Standard closed-form expression used as in [19].

how reviews work

0 comments
Cite this review

Pith. "Pith review of SARFA: Segment Anything with Radiomic Feature Alignment." pith.science (2026). https://pith.science/paper/AKGF5WWL

@misc{pith2026260713323,
  author       = {Pith},
  title        = {Pith review of: SARFA: Segment Anything with Radiomic Feature Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AKGF5WWL}},
  note         = {Machine review of arXiv:2607.13323}
}
read the original abstract

The Segment Anything Model (SAM) has demonstrated strong generalizability across a variety of segmentation tasks. However, SAM often struggles in situations where the target to be segmented is ambiguous. This poses a problem in medical imaging, where accurate delineation of targets such as tumors is vital, but even expert radiologists can disagree on the appropriate boundary for a target. Addressing this, we propose SARFA (Segment Anything with Radiomic Feature Alignment), a novel framework for improved medical image segmentation. Via probabilistic prompting, SARFA generates a diverse set of plausible masks for each input image and optimizes them with a radiomics-driven training objective based on Fr\'echet Radiomic Distance (FRD) and Direct Preference Optimization (DPO). By minimizing the FRD between masked predicted and ground truth regions within each image, SARFA encourages segmentation outputs whose anatomical and textural characteristics align with clinically meaningful ground truth representations, without relying solely on pixel-level overlap. Evaluated on computed tomography (CT) and magnetic resonance imaging (MRI) benchmarks, SARFA outperforms existing ambiguous segmentation methods, demonstrating the effectiveness of radiomic feature alignment and DPO-style candidate mask ranking as a training objective. Our code is available at https://github.com/tbwa233/SARFA.

Figures

Figures reproduced from arXiv: 2607.13323 by the authors.

Figure 1
Figure 1. Overview of the proposed SARFA framework. A probabilistic prompt generator produces diverse prompt embeddings that [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Progression of loss values for P2 SAM and SARFA when trained for 100 epochs on the LIDC-IDRI. Both models exhibit rapid convergence within the first 10 epochs; hence, the selection of this value for the rest of our experiments. Ltotal = Lsup + (λ × LDPO), (17) where Lsup includes cross-entropy, Dice, and Focal losses, and λ is a small weight applied to the DPO loss. To en￾sure stability in the loss calculation, we a… view at source ↗
Figure 3
Figure 3. Visual comparison of selected plausible masks generated [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Visual comparison of selected plausible masks generated [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Correlation analysis between FRD and established segmentation quality metrics across all valid candidate masks generated during [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Visual comparison of plausible masks generated by both P [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Additional visualizations of plausible masks generated by both P [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Visual comparison of plausible masks generated by both P [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Additional visual comparisons of plausible masks generated by both P [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 4 linked inside Pith

  1. [1]

    Samuel G Armato III, Geoffrey McLennan, Luc Bidaut, Michael F McNitt-Gray, Charles R Meyer, Anthony P Reeves, Binsheng Zhao, Denise R Aberle, Claudia I Hen- schke, Eric A Hoffman, et al. The Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI): A completed reference database of lung nodules on CT scans.Medical Physics, 38(2)...

    2011

  2. [2]

    Machine learning and radiomics for segmentation and classification of adnexal masses on ultrasound.NPJ Preci- sion Oncology, 8(1):41, 2024

    Jennifer F Barcroft, Kristofer Linton-Reid, Chiara Lan- dolfo, Maya Al-Memar, Nina Parker, Chris Kyriacou, Maria Munaretto, Martina Fantauzzi, Nina Cooper, Joseph Yazbek, et al. Machine learning and radiomics for segmentation and classification of adnexal masses on ultrasound.NPJ Preci- sion Oncology, 8(1):41, 2024. 3

    2024

  3. [3]

    PHiSeg: Capturing uncertainty in medical im- age segmentation

    Christian F Baumgartner, Kerem C Tezcan, Krishna Chai- tanya, Andreas M H ¨otker, Urs J Muehlematter, Khoschy Schawkat, Anton S Becker, Olivio Donati, and Ender Konukoglu. PHiSeg: Capturing uncertainty in medical im- age segmentation. InInternational Conference on Medi- cal Image Computing and Computer-Assisted Intervention, pages 119–127. Springer, 2019....

    2019

  4. [4]

    GazeRadar: A gaze and radiomics-guided disease localiza- tion framework

    Moinak Bhattacharya, Shubham Jain, and Prateek Prasanna. GazeRadar: A gaze and radiomics-guided disease localiza- tion framework. InInternational Conference on Medical Im- age Computing and Computer-Assisted Intervention, pages 686–696. Springer, 2022. 3

    2022

  5. [5]

    A radiomics-incorporated deep ensemble learning model for multi-parametric MRI-based glioma segmentation.Physics in Medicine & Biology, 68(18):185025, 2023

    Yang Chen, Zhenyu Yang, Jingtong Zhao, Justus Adam- son, Yang Sheng, Fang-Fang Yin, and Chunhao Wang. A radiomics-incorporated deep ensemble learning model for multi-parametric MRI-based glioma segmentation.Physics in Medicine & Biology, 68(18):185025, 2023. 3

    2023

  6. [6]

    From global radiomics to parametric maps: A unified workflow fusing radiomics and deep learning for pdac detection

    Zengtian Deng, Yimeng He, Yu Shi, Lixia Wang, Tou- seef Ahmad Qureshi, Xiuzhen Huang, and Debiao Li. From global radiomics to parametric maps: A unified workflow fusing radiomics and deep learning for pdac detection. In 2026 IEEE 23rd International Symposium on Biomedical Imaging (ISBI), pages 1–5. IEEE, 2026. 3

    2026

  7. [7]

    Graph neural network model using radiomics for lung CT image segmentation.Scientific Reports, 15(1): 34148, 2025

    Mohammad Khalid Faizi, Yan Qiang, Md Masum Billa Sha- gar, Yangyang Wei, Ying Qiao, Juanjuan Zhao, and Zia Urrehman. Graph neural network model using radiomics for lung CT image segmentation.Scientific Reports, 15(1): 34148, 2025. 3

    2025

  8. [8]

    Overview of current biomedical image segmentation methods

    Maedeh Sadat Fasihi and Wasfy B Mikhael. Overview of current biomedical image segmentation methods. In2016 In- ternational Conference on Computational Science and Com- putational Intelligence (CSCI), pages 803–808. IEEE, 2016. 1

    2016

Show all 43 references
  1. [9]

    Medical image seg- mentation: A comprehensive review of deep learning-based methods.Tomography, 11(5):52, 2025

    Yuxiao Gao, Yang Jiang, Yanhong Peng, Fujiang Yuan, Xinyue Zhang, and Jianfeng Wang. Medical image seg- mentation: A comprehensive review of deep learning-based methods.Tomography, 11(5):52, 2025. 1

    2025

  2. [10]

    Modeling multimodal aleatoric uncertainty in segmentation with mixture of stochastic experts

    Zhitong Gao, Yucong Chen, Chuyu Zhang, and Xuming He. Modeling multimodal aleatoric uncertainty in segmentation with mixture of stochastic experts. InThe Eleventh Interna- tional Conference on Learning Representations, 2023. 1, 3, 5, 6

    2023

  3. [11]

    Prostate lesion segmentation in MR images using radiomics based deeply supervised U-Net.Biocybernetics and Biomed- ical Engineering, 40(4):1421–1435, 2020

    Praful Hambarde, Sanjay Talbar, Abhishek Mahajan, Satishkumar Chavan, Meenakshi Thakur, and Nilesh Sable. Prostate lesion segmentation in MR images using radiomics based deeply supervised U-Net.Biocybernetics and Biomed- ical Engineering, 40(4):1421–1435, 2020. 3

    2020

  4. [12]

    Predicting pathological complete response based on weakly and semi-supervised joint learning from breast cancer mri

    Xinyu Hao, Hongming Xu, Nannan Zhao, Tao Yu, Timo Hamalainen, and Fengyu Cong. Predicting pathological complete response based on weakly and semi-supervised joint learning from breast cancer mri. In2023 45th An- nual International Conference of the IEEE Engineering in Medicine...

  5. [13]

    P2SAM: Probabilistically prompted SAMs are ef- ficient segmentator for ambiguous medical images

    Yuzhi Huang, Chenxin Li, Zixu Lin, Hengyu Liu, Haote Xu, Yifan Liu, Yue Huang, Xinghao Ding, Xiaotong Tu, and Yix- uan Yuan. P2SAM: Probabilistically prompted SAMs are ef- ficient segmentator for ambiguous medical images. InPro- ceedings of the 32nd ACM International Conferenc...

    2024

  6. [14]

    Calibrated adversarial refinement for stochastic semantic segmentation

    Elias Kassapis, Georgi Dikov, Deepak K Gupta, and Cedric Nugteren. Calibrated adversarial refinement for stochastic semantic segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 7057– 7067, 2021. 1, 2, 5, 6

    2021

  7. [15]

    Segment any- thing

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer White- head, Alexander C Berg, Wan-Yen Lo, et al. Segment any- thing. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 4015–4026, 2023. 1

    2023

  8. [16]

    A probabilistic u-net for segmentation of ambiguous images

    Simon Kohl, Bernardino Romera-Paredes, Clemens Meyer, Jeffrey De Fauw, Joseph R Ledsam, Klaus Maier-Hein, SM Eslami, Danilo Jimenez Rezende, and Olaf Ronneberger. A probabilistic u-net for segmentation of ambiguous images. Advances in Neural Information Processing Systems, 31,

  9. [17]

    A hierar- chical probabilistic u-net for modeling multi-scale ambigui- ties.arXiv preprint arXiv:1905.13077, 2019

    Simon AA Kohl, Bernardino Romera-Paredes, Klaus H Maier-Hein, Danilo Jimenez Rezende, SM Eslami, Pushmeet Kohli, Andrew Zisserman, and Olaf Ronneberger. A hierar- chical probabilistic u-net for modeling multi-scale ambigui- ties.arXiv preprint arXiv:1905.13077, 2019. 1, 2, 5, 6

    1905 arXiv

  10. [18]

    En- hancing SAM with efficient prompting and preference opti- mization for semi-supervised medical image segmentation

    Aishik Konwer, Zhijian Yang, Erhan Bas, Cao Xiao, Pra- teek Prasanna, Parminder Bhatia, and Taha Kass-Hout. En- hancing SAM with efficient prompting and preference opti- mization for semi-supervised medical image segmentation. InProceedings of the Computer Vision and Pattern R...

    2025

  11. [19]

    Fr\’echet Radiomic 9 Distance (FRD): A versatile metric for comparing medical imaging datasets.arXiv preprint arXiv:2412.01496, 2024

    Nicholas Konz, Richard Osuala, Preeti Verma, Yuwen Chen, Hanxue Gu, Haoyu Dong, Yaqian Chen, Andrew Marshall, Lidia Garrucho, Kaisar Kushibar, et al. Fr\’echet Radiomic 9 Distance (FRD): A versatile metric for comparing medical imaging datasets.arXiv preprint arXiv:2412.01496,...

    2024

  12. [20]

    Deep-radiomic fusion for early detection of pan- creatic ductal adenocarcinoma.Applied Sciences, 15(24): 13024, 2025

    Georgios Lekkas, Eleni Vrochidou, and George A Pa- pakostas. Deep-radiomic fusion for early detection of pan- creatic ductal adenocarcinoma.Applied Sciences, 15(24): 13024, 2025. 3

    2025

  13. [21]

    Generative adversar- ial networks with radiomics supervision for lung lesion gen- eration.IEEE Transactions on Biomedical Engineering, 72 (1):286–296, 2024

    Junyuan Li, Shaoyan Pan, Xiaoxuan Zhang, Cheng Ting Lin, J Webster Stayman, and Grace J Gang. Generative adversar- ial networks with radiomics supervision for lung lesion gen- eration.IEEE Transactions on Biomedical Engineering, 72 (1):286–296, 2024. 3

    2024

  14. [22]

    Segment anything in medical images.Nature Communications, 15(1):654, 2024

    Jun Ma, Yuting He, Feifei Li, Lin Han, Chenyu You, and Bo Wang. Segment anything in medical images.Nature Communications, 15(1):654, 2024. 1

    2024

  15. [23]

    Introduction to radiomics.Journal of Nuclear Medicine, 61 (4):488–495, 2020

    Marius E Mayerhoefer, Andrzej Materka, Georg Langs, Ida H¨aggstr¨om, Piotr Szczypi´nski, Peter Gibbs, and Gary Cook. Introduction to radiomics.Journal of Nuclear Medicine, 61 (4):488–495, 2020. 2

    2020

  16. [24]

    The multimodal brain tumor image segmentation benchmark (BRATS).IEEE Transactions on Medical Imag- ing, 34(10):1993–2024, 2014

    Bjoern H Menze, Andras Jakab, Stefan Bauer, Jayashree Kalpathy-Cramer, Keyvan Farahani, Justin Kirby, Yuliya Burren, Nicole Porz, Johannes Slotboom, Roland Wiest, et al. The multimodal brain tumor image segmentation benchmark (BRATS).IEEE Transactions on Medical Imag- ing, 34(...

    1993

  17. [25]

    Stochastic seg- mentation networks: Modelling spatially correlated aleatoric uncertainty.Advances in Neural Information Processing Sys- tems, 33:12756–12767, 2020

    Miguel Monteiro, Lo ¨ıc Le Folgoc, Daniel Coelho de Castro, Nick Pawlowski, Bernardo Marques, Konstantinos Kamnit- sas, Mark Van der Wilk, and Ben Glocker. Stochastic seg- mentation networks: Modelling spatially correlated aleatoric uncertainty.Advances in Neural Information P...

    2020

  18. [26]

    Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems, 36:53728–53741, 2023

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo- pher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems, 36:53728–53741, 2023. 2, 5

    2023

  19. [27]

    Ambiguous medical image segmentation using diffusion models

    Aimon Rahman, Jeya Maria Jose Valanarasu, Ilker Haci- haliloglu, and Vishal M Patel. Ambiguous medical image segmentation using diffusion models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11536–11546, 2023. 1, 3, 5, 6

    2023

  20. [28]

    AutoSAM: Adapting SAM to medical images by overload- ing the prompt encoder.arXiv preprint arXiv:2306.06370,

    Tal Shaharabany, Aviad Dahan, Raja Giryes, and Lior Wolf. AutoSAM: Adapting SAM to medical images by overload- ing the prompt encoder.arXiv preprint arXiv:2306.06370,

    arXiv

  21. [29]

    Artificial intel- ligence aided diagnosis of pulmonary nodules segmentation and feature extraction.Clinical Radiology, 78(6):437–443,

    T-W Tang, W-Y Lin, J-D Liang, and K-M Li. Artificial intel- ligence aided diagnosis of pulmonary nodules segmentation and feature extraction.Clinical Radiology, 78(6):437–443,

  22. [30]

    Computational radiomics system to de- code the radiographic phenotype.Cancer Research, 77(21): e104–e107, 2017

    Joost JM Van Griethuysen, Andriy Fedorov, Chintan Parmar, Ahmed Hosny, Nicole Aucoin, Vivek Narayan, Regina GH Beets-Tan, Jean-Christophe Fillion-Robin, Steve Pieper, and Hugo JWL Aerts. Computational radiomics system to de- code the radiographic phenotype.Cancer Research, 77(...

    2017

  23. [31]

    Sam2Rad: A seg- mentation model for medical images with learnable prompts

    Assefa Seyoum Wahd, Banafshe Felfeliyan, Yuyue Zhou, Shrimanti Ghosh, Adam McArthur, Jiechen Zhang, Jacob L Jaremko, and Abhilash Hareendranathan. Sam2Rad: A seg- mentation model for medical images with learnable prompts. Computers in Biology and Medicine, 187:109725, 2025. 3

    2025

  24. [32]

    A probabilistic Segment Anything Model for ambiguity-aware medical image seg- mentation

    Tyler Ward and Abdullah Imran. A probabilistic Segment Anything Model for ambiguity-aware medical image seg- mentation. InMedical Imaging 2026: Imaging Informatics, pages 7–12. SPIE, 2026. 1, 3

    2026

  25. [33]

    Annotation- efficient task guidance for medical Segment Anything

    Tyler Ward and Abdullah Al Zubaer Imran. Annotation- efficient task guidance for medical Segment Anything. In 2025 IEEE 22nd International Symposium on Biomedical Imaging (ISBI), pages 1–4. IEEE, 2025. 3

    2025

  26. [34]

    Autoadap- tive medical Segment Anything Model.arXiv preprint arXiv:2507.01828, 2025

    Tyler Ward, Meredith K Owen, O’Kira Coleman, Brian Noehren, and Abdullah-Al-Zubaer Imran. Autoadap- tive medical Segment Anything Model.arXiv preprint arXiv:2507.01828, 2025. 3

    2025

  27. [35]

    Uncertainty as a founda- tion for trustworthy medical imaging AI: A comprehensive review.Authorea Preprints, 2025

    Suyang Xi, Shansong Wang, Mojtaba Safari, Mingzhe Hu, Zhen Tian, and Xiaofeng Yang. Uncertainty as a founda- tion for trustworthy medical imaging AI: A comprehensive review.Authorea Preprints, 2025. 1

    2025

  28. [36]

    MaskSAM: Auto-prompt SAM with mask clas- sification for volumetric medical image segmentation

    Bin Xie, Hao Tang, Bin Duan, Dawen Cai, Yan Yan, and Gady Agam. MaskSAM: Auto-prompt SAM with mask clas- sification for volumetric medical image segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 24423–24433, 2025. 3

    2025

  29. [37]

    Integrating radiomic texture analysis and deep learning for automated myocardial infarc- tion detection in cine-MRI.Scientific Reports, 15(1):24365,

    Wang Xu and Xiangjiang Shi. Integrating radiomic texture analysis and deep learning for automated myocardial infarc- tion detection in cine-MRI.Scientific Reports, 15(1):24365,

  30. [38]

    Deep-learning and radiomics ensemble classifier for false positive reduction in brain metastases segmentation.Physics in Medicine & Biol- ogy, 67(2):025004, 2022

    Zi Yang, Mingli Chen, Mahdieh Kazemimoghadam, Lin Ma, Strahinja Stojadinovic, Robert Timmerman, Tu Dan, Zabi Wardak, Weiguo Lu, and Xuejun Gu. Deep-learning and radiomics ensemble classifier for false positive reduction in brain metastases segmentation.Physics in Medicine & Bi...

    2022

  31. [39]

    Glog-csunet: Enhanc- ing vision transformers with adaptable radiomic features for medical image segmentation

    Niloufar Eghbali Zarch, Hassan Bagher-Ebadian, Tuka Al- hanai, and Mohammad M Ghassemi. Glog-csunet: Enhanc- ing vision transformers with adaptable radiomic features for medical image segmentation. InICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signa...

    2025

  32. [40]

    Customized Segment Any- thing Model for medical image segmentation.arXiv preprint arXiv:2304.13785, 2023

    Kaidong Zhang and Dong Liu. Customized Segment Any- thing Model for medical image segmentation.arXiv preprint arXiv:2304.13785, 2023. 1, 3, 5, 6

    2023 arXiv

  33. [41]

    PixelSeg: Pixel-by-pixel stochastic semantic segmentation for ambiguous medical images

    Wei Zhang, Xiaohong Zhang, Sheng Huang, Yuting Lu, and Kun Wang. PixelSeg: Pixel-by-pixel stochastic semantic segmentation for ambiguous medical images. InProceed- ings of the 30th ACM International Conference on Multime- dia, pages 4742–4750, 2022. 1, 3, 5, 6

    2022

  34. [42]

    SAM-SP: Self-prompting makes SAM great again.arXiv preprint arXiv:2408.12364,

    Chunpeng Zhou, Kangjie Ning, Qianqian Shen, Sheng Zhou, Zhi Yu, and Haishuai Wang. SAM-SP: Self-prompting makes SAM great again.arXiv preprint arXiv:2408.12364,

    arXiv

  35. [43]

    UDNDSNet: A unified deterministic and non-deterministic segmentation network for multi-scene medical image analy- sis.Knowledge-Based Systems, page 114641, 2025

    Haoran Zhou, Ling Xu, Chunmeng Liu, and Guangyao Li. UDNDSNet: A unified deterministic and non-deterministic segmentation network for multi-scene medical image analy- sis.Knowledge-Based Systems, page 114641, 2025. 1 10 A. SARFA Training Algorithm A formal algorithm showing th...

    2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.