pith. sign in

arxiv: 2605.19210 · v1 · pith:PD6YTKDWnew · submitted 2026-05-19 · 💻 cs.CV

D-Convexity: A Unified Differentiable Convex Shape Prior via Quasi-Concavity for Data-driven Image Segmentation

Shengzhe Chen , Hao Yan This is my paper

Pith reviewed 2026-05-20 07:49 UTC · model grok-4.3

classification 💻 cs.CV
keywords convexity priorquasi-concavityimage segmentationdifferentiable shape priorsuper-level setsconvex gradient projectionshape regularity
0
0 comments X

The pith

Requiring quasi-concavity of the continuous mask function enforces convexity on every super-level set for differentiable segmentation.

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

The paper establishes a functional approach to convexity in image segmentation by requiring that all super-level sets of the network's output mask u are convex. This transforms the global shape constraint into local differentiable inequalities on the function and its derivatives at different orders. The resulting losses can be integrated directly into segmentation networks through a convex gradient projection module without needing thresholds. This approach unifies various prior convexity models under one continuous framework and leads to improved shape regularity in practice.

Core claim

The core discovery is that quasi-concavity of u ensures convexity of all its super-level sets, which yields zero-order, first-order, and second-order local characterizations that produce a compact convolutional loss applicable densely across the image. This framework unifies previous discrete and level-set based convex shape priors into a single differentiable setting.

What carries the argument

Quasi-concavity of the mask function u, which converts global convexity requirements into local inequalities on u and its derivatives.

If this is right

  • The convex gradient projection module allows seamless integration with modern segmentation networks.
  • Application of the losses consistently enforces convexity and improves shape regularity on multiple datasets.
  • The method outperforms specialized retinal segmentation networks and prior shape-aware approaches.
  • Previous convex shape models from line constraints to curvature priors are unified within the continuous framework.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Extending the quasi-concavity prior to 3D volumetric segmentation could enforce convex structures in medical imaging.
  • The local inequalities may combine with other priors such as topology preservation for more constrained segmentations.
  • Because the conditions are convolutional and dense, they could support online adaptation in video streams.

Load-bearing premise

Enforcing convexity on all super-level sets of the continuous mask function u is necessary and sufficient for the desired binary segmentation shape while keeping the derivative inequalities stable during optimization.

What would settle it

Observing a trained network producing a non-convex binary segmentation mask despite applying the quasi-concavity loss would falsify the claim that the method reliably enforces convexity.

Figures

Figures reproduced from arXiv: 2605.19210 by Hao Yan, Shengzhe Chen.

Figure 1
Figure 1. Figure 1: Architecture of the proposed framework. ity of a quadratic form along the tangent direction of level sets. The second-order form leads to a local, differen￾tiable penalty that can be computed efficiently within mod￾ern deep learning frameworks and applied densely without relying on any threshold. This penalty is then integrated into the neural network through a convex gradient projec￾tion module (CGPM). Re… view at source ↗
Figure 2
Figure 2. Figure 2: Concave vs. quasi-concave functions. A concave func [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Toy illustration of convexification under the proposed [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative visualization results comparison with shape-aware methods. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Illustration of the first-order condition. If [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Illustration of the second-order condition. For fixed [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
read the original abstract

Convexity is a fundamental geometric prior that underlies many natural and man-made structures, yet remains challenging to impose effectively in end-to-end trainable segmentation networks. We revisit convexity from a functional perspective and propose a unified, threshold-free convexity prior based on the quasi-concavity of the network's output mask function u. Instead of constraining a single binary segmentation, we require all super-level sets of u to be convex, transforming global shape constraints into local, differentiable inequalities on u and its derivatives. From this principle, we derive zero, first, and second-order characterizations, yielding respectively a local midpoint convexification algorithm, a gradient-based condition linked to supporting hyperplanes, and a sufficient second-order inequality expressed as a quadratic form on the tangent plane. The first and second-order formulations produce a compact convolutional loss that can be densely applied across the image without thresholding. Our quasi-concavity losses integrate seamlessly with modern segmentation networks via the proposed convex gradient projection module (CGPM). They consistently enforce convexity and improve shape regularity across multiple datasets, outperforming networks tailored for retinal segmentation and surpassing previous shape-aware methods. Remarkably, our analysis unifies a wide spectrum of previous convex shape models, from discrete 1-0-1 line constraints and graph-cuts convexity formulations to curvature or signed distance Laplacian based level-set priors, within a single continuous and differentiable framework.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces D-Convexity, a unified differentiable convex shape prior for image segmentation networks based on quasi-concavity of the output mask function u. By requiring all super-level sets of u to be convex, it transforms global constraints into local zero-, first-, and second-order differentiable inequalities on u and its derivatives. These are formulated as convolutional losses integrated via the convex gradient projection module (CGPM). The approach is claimed to consistently enforce convexity, improve shape regularity across datasets, outperform specialized networks and previous shape-aware methods, and unify a spectrum of prior convex shape models in a continuous framework.

Significance. If the results hold, this provides a significant contribution by offering a parameter-free, differentiable unification of convexity priors that can be seamlessly incorporated into modern segmentation architectures. The ability to enforce shape regularity without thresholding is particularly valuable for end-to-end training in computer vision applications.

major comments (2)
  1. [Method (characterizations of quasi-concavity)] The derivation of the local derivative inequalities from the quasi-concavity requirement (zero, first, and second-order) is central, but the manuscript does not sufficiently demonstrate that these conditions remain sufficient to guarantee convexity of the binary segmentation after discretization and thresholding on a pixel grid. The stress-test concern that SGD dynamics and finite sampling may permit violations not strongly penalized by the loss is a load-bearing issue for the practical effectiveness of the CGPM.
  2. [Abstract and Experiments] The claim that the proposed losses 'outperform' prior methods and 'surpass previous shape-aware methods' is asserted, but the abstract provides no quantitative tables, ablation studies, or error bars; if these are present in the full manuscript, they should be referenced explicitly to support the empirical superiority.
minor comments (2)
  1. Clarify the exact definition of the convex gradient projection module (CGPM) and how it differs from standard projection methods in the introduction or early method section.
  2. [Notation] Ensure consistent use of symbols for the mask function u and the level t across the paper.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and insightful comments on our manuscript. These points help strengthen the presentation of the quasi-concavity framework and the empirical claims. We address each major comment below and indicate the revisions we will make.

read point-by-point responses

  1. Referee: [Method (characterizations of quasi-concavity)] The derivation of the local derivative inequalities from the quasi-concavity requirement (zero, first, and second-order) is central, but the manuscript does not sufficiently demonstrate that these conditions remain sufficient to guarantee convexity of the binary segmentation after discretization and thresholding on a pixel grid. The stress-test concern that SGD dynamics and finite sampling may permit violations not strongly penalized by the loss is a load-bearing issue for the practical effectiveness of the CGPM.

    Authors: We thank the referee for this important observation. The zero-, first-, and second-order conditions are derived directly from the definition of quasi-concavity, which is equivalent to convexity of every super-level set of the continuous function u. Because the binary segmentation corresponds to a super-level set (for any threshold), convexity is preserved by construction in the continuous case; the threshold-free design avoids post-hoc binarization issues. On a discrete pixel grid the convolutional losses provide a dense approximation to these local inequalities. We agree that the manuscript would benefit from explicit discussion of discretization effects and empirical validation under SGD. In the revision we will add a dedicated subsection with discretization analysis and stress-test experiments (including quantitative convexity violation rates before/after thresholding across training trajectories) to demonstrate that the CGPM loss strongly penalizes violations in practice. revision: yes

  2. Referee: [Abstract and Experiments] The claim that the proposed losses 'outperform' prior methods and 'surpass previous shape-aware methods' is asserted, but the abstract provides no quantitative tables, ablation studies, or error bars; if these are present in the full manuscript, they should be referenced explicitly to support the empirical superiority.

    Authors: The referee correctly notes that the abstract would be strengthened by explicit pointers to the supporting evidence. The full manuscript already contains the requested material: quantitative tables comparing Dice/IoU scores against specialized retinal networks and prior shape-aware methods, ablation studies isolating the contribution of each order of the quasi-concavity loss, and error bars from repeated runs with statistical significance (Sections 4.2–4.4 and supplementary material). We will revise the abstract to reference these results directly (e.g., “outperforming prior methods by up to X% Dice on dataset Y, see Table Z and ablation in Section 4.3”) while preserving its concise style. revision: yes

Circularity Check

0 steps flagged

No circularity: derivation starts from external quasi-concavity definition and produces independent local losses

full rationale

The paper begins from the standard mathematical definition of quasi-concavity (super-level sets convex for all thresholds), which is an external property independent of the paper's own outputs or fitted parameters. It then derives zero-, first-, and second-order local conditions on u and its derivatives to produce convolutional losses and the CGPM module. The unification claim is presented as an analytical consequence of this framework rather than a renaming or self-referential reduction. No equations in the provided text show a fitted input being relabeled as a prediction, no load-bearing self-citation chains, and no ansatz smuggled via prior work by the same authors. The central claims remain self-contained against external benchmarks such as standard convexity definitions and prior shape priors.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the mathematical premise that quasi-concavity of the mask function is an appropriate continuous relaxation of binary convexity, together with the modeling choice that local derivative inequalities can be stably optimized inside a neural network.

axioms (1)
  • domain assumption All super-level sets of the network output u must be convex to enforce the desired shape prior.
    This premise is invoked when the authors transform global shape constraints into local inequalities on u and its derivatives.

pith-pipeline@v0.9.0 · 5777 in / 1311 out tokens · 26569 ms · 2026-05-20T07:49:47.724486+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages

  1. [1]

    Quasi-concave pro- gramming.Econometrica: Journal of the Econometric Soci- ety, pages 779–800, 1961

    Kenneth J Arrow and Alain C Enthoven. Quasi-concave pro- gramming.Econometrica: Journal of the Econometric Soci- ety, pages 779–800, 1961. 1, 3

    work page 1961

  2. [2]

    Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: Is the problem solved?IEEE Transactions on Medical Imaging, 37(11): 2514–2525, 2018

    Olivier Bernard, Alain Lalande, Clement Zotti, Frederic Cer- venansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: Is the problem solved?IEEE Transactions on Medical Imaging, 37(11): 2514–2525, 2018. 7

    work page 2018

  3. [3]

    Swin-unet: Unet-like pure transformer for medical image segmentation

    Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xi- aopeng Zhang, Qi Tian, and Manning Wang. Swin-unet: Unet-like pure transformer for medical image segmentation. InEuropean Conference on Computer Vision, pages 205– 218, 2022. 7, 8

    work page 2022

  4. [4]

    Dcan: deep contour-aware networks for accurate gland seg- mentation

    Hao Chen, Xiaojuan Qi, Lequan Yu, and Pheng-Ann Heng. Dcan: deep contour-aware networks for accurate gland seg- mentation. InProceedings of the IEEE Conference on Com- puter Vision and Pattern Recognition, pages 2487–2496,

    work page

  5. [5]

    Encoder-decoder with atrous separable convolution for semantic image segmentation

    Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of European Conference on Computer Vision, pages 833–851, 2018. 8

    work page 2018

  6. [6]

    Contour flow constraint: Preserving global shape similarity for deep learning-based image segmentation.IEEE Transactions on Image Processing, 34:5054–5067, 2025

    Shengzhe Chen, Zhaoxuan Dong, and Jun Liu. Contour flow constraint: Preserving global shape similarity for deep learning-based image segmentation.IEEE Transactions on Image Processing, 34:5054–5067, 2025. 1, 6

    work page 2025

  7. [7]

    Joint optic disc and cup seg- mentation based on multi-label deep network and polar trans- formation.IEEE Transactions on Medical Imaging, 37(7): 1597–1605, 2018

    Huazhu Fu, Jun Cheng, Yanwu Xu, Damon Wing Kee Wong, Jiang Liu, and Xiaochun Cao. Joint optic disc and cup seg- mentation based on multi-label deep network and polar trans- formation.IEEE Transactions on Medical Imaging, 37(7): 1597–1605, 2018. 7

    work page 2018

  8. [8]

    Interactive tool and database for optic disc and cup segmentation of stereo and monocular retinal fundus images

    Francisco Fumero, Jose Sigut, Silvia Alay ´on, M Gonz ´alez- Hern´andez, and M Gonz´alez de la Rosa. Interactive tool and database for optic disc and cup segmentation of stereo and monocular retinal fundus images. 2015. 7

    work page 2015

  9. [9]

    Convexity shape prior for segmentation

    Lena Gorelick, Olga Veksler, Yuri Boykov, and Claudia Nieuwenhuis. Convexity shape prior for segmentation. In European Conference on Computer Vision, pages 675–690. Springer, 2014. 2

    work page 2014

  10. [10]

    Convexity shape prior for binary segmenta- tion.IEEE Transactions on Pattern Analysis & Machine In- telligence, 39(02):258–271, 2017

    Lena Gorelick, Olga Veksler, Yuri Boykov, and Claudia Nieuwenhuis. Convexity shape prior for binary segmenta- tion.IEEE Transactions on Pattern Analysis & Machine In- telligence, 39(02):258–271, 2017. 1

    work page 2017

  11. [11]

    Ce-net: Context encoder network for 2d medical image segmentation.IEEE Transactions on Medical Imag- ing, 38(10):2281–2292, 2019

    Zaiwang Gu, Jun Cheng, Huazhu Fu, Kang Zhou, Huay- ing Hao, Yitian Zhao, Tianyang Zhang, Shenghua Gao, and Jiang Liu. Ce-net: Context encoder network for 2d medical image segmentation.IEEE Transactions on Medical Imag- ing, 38(10):2281–2292, 2019. 7

    work page 2019

  12. [12]

    Geodesic star convexity for interactive image segmentation

    Varun Gulshan, Carsten Rother, Antonio Criminisi, Andrew Blake, and Andrew Zisserman. Geodesic star convexity for interactive image segmentation. In2010 IEEE Computer So- ciety Conference on Computer Vision and Pattern Recogni- tion, pages 3129–3136, 2010. 2

    work page 2010

  13. [13]

    Noise-robust pupil center detection through cnn-based segmentation with shape-prior loss.IEEE Access, 8:64739– 64749, 2020

    Sang Yoon Han, Hyuk Jin Kwon, Yoonsik Kim, and Nam Ik Cho. Noise-robust pupil center detection through cnn-based segmentation with shape-prior loss.IEEE Access, 8:64739– 64749, 2020. 2, 3, 5

    work page 2020

  14. [14]

    Hedgehog shape priors for multi-object segmentation

    Hossam Isack, Olga Veksler, Milan Sonka, and Yuri Boykov. Hedgehog shape priors for multi-object segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2434–2442, 2016. 2

    work page 2016

  15. [15]

    Statistical shape influence in geodesic active con- tours

    Michael E Leventon, W Eric L Grimson, and Olivier Faugeras. Statistical shape influence in geodesic active con- tours. InProceedings of IEEE Conference on Computer Vi- sion and Pattern Recognition, pages 316–323, 2000. 1

    work page 2000

  16. [16]

    Odcs-nsnp: Optic disc and cup segmentation us- ing deep networks enhanced by nonlinear spiking neural p systems.Biomedical Signal Processing and Control, 108: 107935, 2025

    Wang Li, Meichen Xia, Hong Peng, Zhicai Liu, and Jun Guo. Odcs-nsnp: Optic disc and cup segmentation us- ing deep networks enhanced by nonlinear spiking neural p systems.Biomedical Signal Processing and Control, 108: 107935, 2025. 7

    work page 2025

  17. [17]

    Convex shape prior for deep neural convolution network based eye fun- dus images segmentation.arXiv preprint arXiv:2005.07476,

    Jun Liu, Xue-Cheng Tai, and Shousheng Luo. Convex shape prior for deep neural convolution network based eye fun- dus images segmentation.arXiv preprint arXiv:2005.07476,

    work page arXiv 2005

  18. [18]

    Jun Liu, Xiangyue Wang, and Xue-Cheng Tai. Deep con- volutional neural networks with spatial regularization, vol- ume and star-shape priors for image segmentation.Journal of Mathematical Imaging and Vision, 64(6):625–645, 2022. 1

    work page 2022

  19. [19]

    An end-to-end real-time lightweight network for the joint segmentation of optic disc and optic cup on fundus images.Mathematics, 10(22):4288, 2022

    Zhijie Liu, Yuanqiong Chen, Xiaohua Xiang, Zhan Li, Bolin Liao, and Jianfeng Li. An end-to-end real-time lightweight network for the joint segmentation of optic disc and optic cup on fundus images.Mathematics, 10(22):4288, 2022. 7

    work page 2022

  20. [20]

    Convex shape prior for multi-object segmentation using a single level set function

    Shousheng Luo, Xue-Cheng Tai, Limei Huo, Yang Wang, and Roland Glowinski. Convex shape prior for multi-object segmentation using a single level set function. InProceed- ings of the IEEE/CVF International Conference on Com- puter Vision, pages 613–621, 2019. 2, 5, 6

    work page 2019

  21. [21]

    Convex object (s) characterization and segmentation using level set function.Journal of Mathematical Imaging and Vi- sion, 64(1):68–88, 2022

    Shousheng Luo, Xue-Cheng Tai, and Roland Glowinski. Convex object (s) characterization and segmentation using level set function.Journal of Mathematical Imaging and Vi- sion, 64(1):68–88, 2022. 1

    work page 2022

  22. [22]

    A new binary representation method for shape convexity and appli- cation to image segmentation.Analysis & Applications, 20 (3), 2022

    Shousheng Luo, Xue-Cheng Tai, and Yang Wang. A new binary representation method for shape convexity and appli- cation to image segmentation.Analysis & Applications, 20 (3), 2022. 2, 5

    work page 2022

  23. [23]

    A binary characterization method for shape convex- ity and applications.Applied Mathematical Modelling, 122: 780–795, 2023

    Shousheng Luo, Jinfeng Chen, Yunhai Xiao, and Xue-Cheng Tai. A binary characterization method for shape convex- ity and applications.Applied Mathematical Modelling, 122: 780–795, 2023. 2, 5

    work page 2023

  24. [24]

    Deep retinal image understanding

    Kevis-Kokitsi Maninis, Jordi Pont-Tuset, Pablo Arbel ´aez, and Luc Van Gool. Deep retinal image understanding. In International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 140–148. Springer, 2016. 7

    work page 2016

  25. [25]

    Conv-mcd: A plug-and-play multi-task module for medical image segmentation

    Balamurali Murugesan, Kaushik Sarveswaran, Sharath M Shankaranarayana, Keerthi Ram, Jayaraj Joseph, and Mo- hanasankar Sivaprakasam. Conv-mcd: A plug-and-play multi-task module for medical image segmentation. InInter- national Workshop on Machine Learning in Medical Imag- ing, pages 292–300. Springer, 2019. 7 9

    work page 2019

  26. [26]

    Jos ´e Ignacio Orlando, Huazhu Fu, Jo ˜ao Barbosa Breda, Karel van Keer, Deepti R. Bathula, Andr ´es Diaz-Pinto, Ruogu Fang, Pheng-Ann Heng, Jeyoung Kim, JoonHo Lee, Joonseok Lee, Xiaoxiao Li, Peng Liu, Shuai Lu, Balamu- rali Murugesan, Valery Naranjo, Sai Samarth R. Phaye, Sharath M. Shankaranarayana, Apoorva Sikka, Jaemin Son, Anton van den Hengel, Shuj...

    work page 2020

  27. [27]

    Convex segments for con- vex objects using dnn boundary tracing and graduated op- timization

    Jimut B Pal and Suyash P Awate. Convex segments for con- vex objects using dnn boundary tracing and graduated op- timization. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 91–101. Springer, 2024. 3

    work page 2024

  28. [28]

    U-net: Convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI), pages 234–

    work page

  29. [29]

    Springer, 2015. 7, 8

    work page 2015

  30. [30]

    Convexity shape con- straints for image segmentation

    Loic A Royer, David L Richmond, Carsten Rother, Bjoern Andres, and Dagmar Kainmueller. Convexity shape con- straints for image segmentation. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 402–410, 2016. 1, 2, 3

    work page 2016

  31. [31]

    Cded-net: Joint segmentation of optic disc and optic cup for glaucoma screening.IEEE Access, 8:102733–102747, 2020

    Munazza Tabassum, Tariq M Khan, Muhammad Arsalan, Syed Saud Naqvi, Mansoor Ahmed, Hussain Ahmed Madni, and Jawad Mirza. Cded-net: Joint segmentation of optic disc and optic cup for glaucoma screening.IEEE Access, 8:102733–102747, 2020. 7

    work page 2020

  32. [32]

    Deep multi-task and task- specific feature learning network for robust shape preserved organ segmentation

    Chaowei Tan, Liang Zhao, Zhennan Yan, Kang Li, Dim- itris Metaxas, and Yiqiang Zhan. Deep multi-task and task- specific feature learning network for robust shape preserved organ segmentation. In2018 IEEE 15th International Sym- posium on Biomedical Imaging (ISBI 2018), pages 1221– 1224, 2018. 7

    work page 2018

  33. [33]

    Efficient convex optimization-based curva- ture dependent contour evolution approach for medical im- age segmentation

    Eranga Ukwatta, Jing Yuan, Wu Qiu, Martin Rajchl, and Aaron Fenster. Efficient convex optimization-based curva- ture dependent contour evolution approach for medical im- age segmentation. InMedical Imaging 2013: Image Pro- cessing, page 866902. SPIE, 2013. 2, 5

    work page 2013

  34. [34]

    Star shape prior for graph-cut image segmen- tation

    Olga Veksler. Star shape prior for graph-cut image segmen- tation. InEuropean Conference on Computer Vision, pages 454–467. Springer, 2008. 2

    work page 2008

  35. [35]

    Caiyong Wang, Jawad Muhammad, Yunlong Wang, Zhaofeng He, and Zhenan Sun. Towards complete and ac- curate iris segmentation using deep multi-task attention net- work for non-cooperative iris recognition.IEEE Transac- tions on Information Forensics and Security, 15:2944–2959,

    work page

  36. [36]

    Active boundary loss for semantic segmentation

    Chi Wang, Yunke Zhang, Miaomiao Cui, Peiran Ren, Yin Yang, Xuansong Xie, Xian-Sheng Hua, Hujun Bao, and Wei- wei Xu. Active boundary loss for semantic segmentation. InProceedings of the AAAI Conference on Artificial Intelli- gence, pages 2397–2405, 2022. 7

    work page 2022

  37. [37]

    Towards an ex- tended efficientnet-based u-net framework for joint optic disc and cup segmentation in the fundus image.Biomedical Sig- nal Processing and Control, 85:104906, 2023

    Jinke Wang, Xiang Li, and Yuanzhi Cheng. Towards an ex- tended efficientnet-based u-net framework for joint optic disc and cup segmentation in the fundus image.Biomedical Sig- nal Processing and Control, 85:104906, 2023. 7

    work page 2023

  38. [38]

    Mixed maximum loss design for optic disc and optic cup segmenta- tion with deep learning from imbalanced samples.Sensors, 19(20):4401, 2019

    Yong-li Xu, Shuai Lu, Han-xiong Li, and Rui-rui Li. Mixed maximum loss design for optic disc and optic cup segmenta- tion with deep learning from imbalanced samples.Sensors, 19(20):4401, 2019. 7

    work page 2019

  39. [39]

    Convexity shape prior for level set-based image segmenta- tion method.IEEE Transactions on Image Processing, 29: 7141–7152, 2020

    Shi Yan, Xue-Cheng Tai, Jun Liu, and Hai-Yang Huang. Convexity shape prior for level set-based image segmenta- tion method.IEEE Transactions on Image Processing, 29: 7141–7152, 2020. 1, 2, 5

    work page 2020

  40. [40]

    A level set method for convexity preserving segmentation of cardiac left ventricle

    Cong Yang, Xue Shi, Donglan Yao, and Chunming Li. A level set method for convexity preserving segmentation of cardiac left ventricle. In2017 IEEE International Confer- ence on Image Processing (ICIP), pages 2159–2163, 2017. 2, 5

    work page 2017

  41. [41]

    Convex combination star shape prior for data-driven image semantic segmentation

    Xinyu Zhao, Jun Xie, Shengzhe Chen, and Jun Liu. Convex combination star shape prior for data-driven image semantic segmentation. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 14068–14077, 2025. 1

    work page 2025

  42. [42]

    Gdcseg-net: general op- tic disc and cup segmentation network for multi-device fun- dus images.Biomedical Optics Express, 12(10):6529–6544,

    Qianlong Zhu, Xinjian Chen, Qingquan Meng, Jiahuan Song, Gaohui Luo, Meng Wang, Fei Shi, Zhongyue Chen, Dehui Xiang, Lingjiao Pan, et al. Gdcseg-net: general op- tic disc and cup segmentation network for multi-device fun- dus images.Biomedical Optics Express, 12(10):6529–6544,

    work page

  43. [43]

    Proof of Theorem 1 Zero-order quasi-concavity condition:u∈C 0 is quasi- concave⇐ ⇒For anyx,y∈Ω, λ∈[0,1],u(λx+ (1− λ)y)≥min{u(x), u(y)}

    7 10 D-Convexity: A Unified Differentiable Convex Shape Prior via Quasi-Concavity for Data-driven Image Segmentation Supplementary Material A. Proof of Theorem 1 Zero-order quasi-concavity condition:u∈C 0 is quasi- concave⇐ ⇒For anyx,y∈Ω, λ∈[0,1],u(λx+ (1− λ)y)≥min{u(x), u(y)}. Proof.(⇒) Denoteγ= min{u(x), u(y)}. Apparently u(x), u(y)≥γ, which meansx,y∈S ...

    work page