arxiv: 2602.05880 · v2 · submitted 2026-02-05 · 💻 cs.CV

Recognition: 2 theorem links

· Lean Theorem

Contour Refinement using Discrete Diffusion in Low Data Regime

Fei Yu Guan , Ian Keefe , Sophie Wilkinson , Daniel D.B. Perrakis , Steven Waslander

Authors on Pith no claims yet

Pith reviewed 2026-05-16 06:51 UTC · model grok-4.3

classification 💻 cs.CV

keywords contour refinementdiscrete diffusionlow data regimeboundary detectionimage segmentationCNN self-attentionmedical imaging

0 comments

The pith

A lightweight discrete diffusion pipeline refines sparse contours into accurate boundaries using a CNN with self-attention conditioned on segmentation masks.

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

The paper develops a method for detecting precise object boundaries in images when only small amounts of labeled data are available. It begins with a rough segmentation mask and applies iterative denoising to a sparse contour representation inside a simplified diffusion process. A customized convolutional network with self-attention layers serves as the denoiser and produces dense isolated contours with little extra processing. This design targets irregular and translucent objects in settings such as medical imaging and environmental monitoring where training sets contain fewer than 500 examples. The approach yields higher accuracy than several existing methods on one medical dataset, matches them on two others, and runs inference 3.5 times faster.

Core claim

The authors show that a simplified discrete diffusion process, paired with a CNN architecture that includes self-attention layers and conditioned on an input segmentation mask, can iteratively denoise a sparse contour representation to yield a dense isolated boundary. The pipeline incorporates a streamlined diffusion schedule, a tailored network design, and minimal post-processing steps so that the entire refinement operates reliably on datasets smaller than 500 training images.

What carries the argument

The discrete diffusion contour refinement pipeline that iteratively denoises a sparse contour representation conditioned on a segmentation mask inside a customized CNN with self-attention layers.

If this is right

Boundary detection becomes feasible for translucent objects in medical scans without collecting large annotated sets.
Inference speed increases by a factor of 3.5 relative to prior contour methods while maintaining competitive accuracy.
The same pipeline applies directly to environmental tasks such as delineating smoke plumes in wildfire imagery.
Minimal post-processing requirements allow the output contours to be used immediately in downstream measurement or tracking systems.

Where Pith is reading between the lines

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

The approach may reduce the annotation burden in any domain where precise edges matter more than full masks, such as industrial inspection of translucent parts.
Because the diffusion steps are simplified, the method could be further accelerated or quantized for deployment on devices with tight memory limits.
Testing the pipeline on video sequences would reveal whether temporal consistency emerges naturally from the per-frame contour refinement.

Load-bearing premise

The simplified diffusion schedule together with the custom CNN and self-attention layers will produce stable contour refinements across different low-data domains without overfitting or needing extensive extra tuning.

What would settle it

A clear drop in boundary accuracy or an increase in fragmented contours when the same pipeline is evaluated on a new dataset of fewer than 100 images drawn from a previously unseen object class would falsify the claim of robust low-data generalization.

Figures

Figures reproduced from arXiv: 2602.05880 by Daniel D.B. Perrakis, Fei Yu Guan, Ian Keefe, Sophie Wilkinson, Steven Waslander.

**Figure 2.** Figure 2: An overview of the network architecture, training and inference pipelines. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Figure 3: An illustration of the boundary truncation mask. The output from our [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Examples drawn from HAM10K. In this figure and following figures, the top row represents unrefined contours, the middle row represents refined [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Examples drawn from smoke dataset. The top row rep- resents unrefined contours, the middle row represents refined and the last row represents [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

**Figure 6.** Figure 6: Examples drawn from KVASIR. The top row rep- resents unrefined contours, the middle row represents refined and the last row represents ground [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗

read the original abstract

Boundary detection of irregular and translucent objects is an important problem with applications in medical imaging, environmental monitoring and manufacturing, where many of these applications are plagued with scarce labeled data and low in situ computational resources. While recent image segmentation studies focus on segmentation mask alignment with ground-truth, the task of boundary detection remains understudied, especially in the low data regime. In this work, we present a lightweight discrete diffusion contour refinement pipeline for robust boundary detection in the low data regime. We use a Convolutional Neural Network(CNN) architecture with self-attention layers as the core of our pipeline, and condition on a segmentation mask, iteratively denoising a sparse contour representation. We introduce multiple novel adaptations for improved low-data efficacy and inference efficiency, including using a simplified diffusion process, a customized model architecture, and minimal post processing to produce a dense, isolated contour given a dataset of size <500 training images. Our method outperforms several SOTA baselines on the medical imaging dataset KVASIR, is competitive on HAM10K and our custom wildfire dataset, Smoke, while improving inference framerate by 3.5X.

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.

Desk Editor's Note private letter to a colleague

This paper adapts discrete diffusion for contour refinement in low-data settings with practical tweaks, but its outperformance claims need the full experimental details to hold up.

read the letter

The punchline is that the authors have built a specialized discrete diffusion pipeline for refining object boundaries from coarse masks, aimed at situations with very limited training data. They simplify the diffusion steps, use a custom CNN with self-attention, and minimize post-processing to produce clean contours quickly from under 500 images. What the paper does well is target an understudied corner: boundary detection for irregular or translucent objects in medical and environmental images, where full segmentation might not be the end goal. The engineering choices for low-data efficacy and inference speed (claimed 3.5X improvement) address real constraints in those domains. The pipeline seems straightforward to implement based on the description. The soft spots are in the validation. The abstract claims better results than SOTA on KVASIR, competitive on HAM10K and their Smoke wildfire set, but supplies no actual metrics, dataset breakdowns, or statistical tests. Without ablations showing how the simplifications perform as data shrinks, or direct comparisons to standard diffusion baselines, it's difficult to know if the gains come from the method or from dataset-specific fitting. The low-data robustness is asserted rather than demonstrated in the provided summary. This work is for computer vision practitioners focused on segmentation and boundary tasks in data-poor environments, such as medical imaging or remote sensing. A reader looking for new diffusion variants in CV would get some value from the adaptations described, provided the experiments check out. I would bring it to a reading group as a maybe, to discuss the low-data adaptations specifically. I wouldn't cite it yet without seeing the numbers. It deserves peer review because the problem is relevant and the approach is novel enough in its combination to warrant referee input on the details.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes a lightweight discrete diffusion contour refinement pipeline for boundary detection of irregular objects in low-data regimes (<500 training images). It uses a CNN with self-attention layers conditioned on a segmentation mask to iteratively denoise a sparse contour representation, with adaptations including a simplified diffusion process, customized architecture, and minimal post-processing. The central claim is that this yields outperformance versus SOTA baselines on the KVASIR medical dataset, competitiveness on HAM10K and the custom Smoke wildfire dataset, and 3.5X faster inference.

Significance. If the performance and efficiency claims hold under scrutiny, the approach could address a practical gap in boundary detection for data-scarce domains such as medical imaging and environmental monitoring, where both accuracy and low-resource inference matter. The focus on simplified diffusion for low-data efficacy is a potentially useful engineering direction if supported by ablations.

major comments (3)

[Abstract] Abstract: The claims of outperformance on KVASIR, competitiveness on HAM10K/Smoke, and 3.5X inference speedup are stated without any quantitative metrics, error bars, dataset sizes, baseline tables, or statistical tests, making it impossible to assess whether the central claim is supported by evidence.
[§4 (Experiments)] §4 (Experiments): No ablation varying training-set cardinality is reported, nor any comparison against a standard (non-simplified) diffusion baseline; this leaves the low-data generalization assumption untested and the robustness to data scarcity unverified.
[§3 (Method)] §3 (Method): The simplified diffusion process and self-attention CNN are presented as key for low-data efficacy, but without explicit equations showing the schedule or conditioning, or controls for overfitting, the load-bearing claim that these adaptations suffice for <500 images remains unsubstantiated.

minor comments (2)

[§3 (Method)] The notation for the sparse contour representation and denoising steps should be formalized with equations to improve reproducibility.
[§4 (Experiments)] Figure captions and table headers lack sufficient detail on metric definitions (e.g., which boundary-specific IoU or F-score variant is used).

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their valuable feedback. We believe the suggested revisions will strengthen the presentation of our work. We address each major comment below, indicating the changes made to the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: The claims of outperformance on KVASIR, competitiveness on HAM10K/Smoke, and 3.5X inference speedup are stated without any quantitative metrics, error bars, dataset sizes, baseline tables, or statistical tests, making it impossible to assess whether the central claim is supported by evidence.

Authors: We agree with this observation. To address this, we have revised the abstract to include key quantitative metrics from our experiments, such as the performance scores on KVASIR, dataset sizes, and the inference speedup, along with references to the relevant tables and figures. Error bars and details on statistical tests are now explicitly mentioned in the abstract and elaborated in Section 4. revision: yes
Referee: [§4 (Experiments)] §4 (Experiments): No ablation varying training-set cardinality is reported, nor any comparison against a standard (non-simplified) diffusion baseline; this leaves the low-data generalization assumption untested and the robustness to data scarcity unverified.

Authors: We acknowledge this limitation in the original submission. In the revised manuscript, we have included an ablation study that varies the training set cardinality (from 100 to the full <500 images) to verify the low-data generalization. We have also added a comparison against a standard (non-simplified) discrete diffusion baseline, which demonstrates the benefits of our simplifications in terms of both accuracy and efficiency in data-scarce settings. revision: yes
Referee: [§3 (Method)] §3 (Method): The simplified diffusion process and self-attention CNN are presented as key for low-data efficacy, but without explicit equations showing the schedule or conditioning, or controls for overfitting, the load-bearing claim that these adaptations suffice for <500 images remains unsubstantiated.

Authors: We have updated Section 3 with explicit equations for the simplified diffusion schedule and the conditioning mechanism. The forward and reverse processes are now formalized mathematically. Furthermore, we have included details on the self-attention integration and added controls for overfitting, such as the use of dropout layers, L2 regularization, and monitoring of validation loss to prevent overfitting on small datasets. revision: yes

Circularity Check

0 steps flagged

No circularity detected; empirical pipeline is self-contained engineering contribution

full rationale

The paper describes an applied method: a CNN with self-attention layers that iteratively denoises a sparse contour representation conditioned on a segmentation mask, using a simplified diffusion process and minimal post-processing. No equations or derivations are presented that reduce a claimed result to a fitted parameter or self-referential definition. Performance claims (outperformance on KVASIR, competitiveness on HAM10K/Smoke, 3.5X faster inference) rest on reported empirical evaluations rather than any quantity defined by construction from the inputs. No self-citation chains, uniqueness theorems, or ansatzes are invoked in a load-bearing way. The derivation chain consists of independent architectural and process choices for low-data regimes and is therefore self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Review performed on abstract only; no explicit free parameters, axioms, or invented entities are stated. The approach builds on standard CNN and diffusion concepts with task-specific simplifications whose details are not provided.

pith-pipeline@v0.9.0 · 5499 in / 1098 out tokens · 33040 ms · 2026-05-16T06:51:15.828527+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We use a Convolutional Neural Network(CNN) architecture with self-attention layers as the core of our pipeline, and condition on a segmentation mask, iteratively denoising a sparse contour representation. We introduce multiple novel adaptations for improved low-data efficacy and inference efficiency, including using a simplified diffusion process
IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Our losses consist of only the simple loss component (eq. (5)) , as in [14], since we find training with the full KL matching loss requires extensive amounts of data... In the low data setting, the DICE Loss is used instead of the full KL matching loss

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · 13 internal anchors

[1]

SAM 2: Segment Anything in Images and Videos, Nikhila Ravi and Valentin Gabeur and Yuan-Ting Hu and Ronghang Hu and Chaitanya Ryali and Tengyu Ma and Haitham Khedr and Roman R ¨adle and Chloe Rolland and Laura Gustafson and Eric Mintun and Junting Pan and Kalyan Vasudev Alwala and Nicolas Carion and Chao-Yuan Wu and Ross Girshick and Piotr Doll ´ar and Ch...

work page internal anchor Pith review Pith/arXiv arXiv 2024
[2]

SegRefiner: Towards Model-Agnostic Segmentation Refinement with Discrete Diffusion Process, Mengyu Wang and Henghui Ding and Jun Hao Liew and Jiajun Liu and Yao Zhao and Yunchao Wei, 2023, 2312.12425, arXiv, cs.CV, https://arxiv.org/abs/2312.12425,

work page arXiv 2023
[3]

Mazurowski, 2024, 2403.10786, arXiv, eess.IV, https://arxiv.org/abs/2403.10786,

ContourDiff: Unpaired Image-to-Image Translation with Structural Con- sistency for Medical Imaging, Yuwen Chen and Nicholas Konz and Hanxue Gu and Haoyu Dong and Yaqian Chen and Lin Li and Jisoo Lee and Maciej A. Mazurowski, 2024, 2403.10786, arXiv, eess.IV, https://arxiv.org/abs/2403.10786,

work page arXiv 2024
[4]

Wei-Ying Ma and Manjunath, B.S., IEEE Transactions on Image Processing, EdgeFlow: a technique for boundary detection and image segmentation, 2000, 1375-1388, Image edge detection;Image segmentation;Predictive coding;Predictive models;Computer vision;Image retrieval;Content based retrieval;Gabor filters;Filtering;Application software, 10.1109/83.855433

work page doi:10.1109/83.855433 2000
[5]

and Kimmel, R

Caselles, V . and Kimmel, R. and Sapiro, G., bookProceedings of IEEE International Conference on Computer Vision, Geodesic active contours, 1995, 694-699, 10.1109/ICCV .1995.466871

work page doi:10.1109/iccv 1995
[6]

MedSegDiff: Medical Image Segmentation with Diffusion Probabilistic Model, Junde Wu and Rao Fu and Huihui Fang and Yu Zhang and Yehui Yang and Haoyi Xiong and Huiying Liu and Yanwu Xu, 2023, 2211.00611, arXiv, cs.CV, https://arxiv.org/abs/2211.00611,

work page arXiv 2023
[8]

Deep Smoke Segmentation, Feiniu Yuan and Lin Zhang and Xue Xia and Boyang Wan and Qinghua Huang and Xuelong Li, 2018, 1809.00774, arXiv, cs.CV, https://arxiv.org/abs/1809.00774,

work page internal anchor Pith review Pith/arXiv arXiv 2018
[9]

Li, Mengna and Zhang, Youmin and Mu, Lingxia and Jing, Xin and Yu, Ziquan and Jiao, Shangbin and Liu, Han and Xie, Guo and Yingmin, Yi, 2022, 01, 145-150, A Real-time Fire Segmentation Method Based on A Deep Learning Approach, 55, IFAC-PapersOnLine, 10.1016/j.ifacol.2022.07.120

work page doi:10.1016/j.ifacol.2022.07.120 2022
[10]

Rethinking Atrous Convolution for Semantic Image Segmenta- tion, Liang-Chieh Chen and George Papandreou and Florian Schroff and Hartwig Adam, 2017, 1706.05587, arXiv, cs.CV, https://arxiv.org/abs/1706.05587,

work page internal anchor Pith review Pith/arXiv arXiv 2017
[11]

Using DUCK-Net for polyp image segmentation, 13, 2045-2322, http://dx.doi.org/10.1038/s41598-023-36940-5, 10.1038/s41598-023- 36940-5, 1, Scientific Reports, publisher=Springer Science and Business Media LLC, Dumitru, Razvan-Gabriel and Peteleaza, Darius and Craciun, Catalin, 2023, month=jun

work page doi:10.1038/s41598-023-36940-5 2045
[12]

Unpublished, Glenn Jocher and Jing Qiu, Ultralytics YOLO11, version = 11.0.0, 2024, https://github.com/ultralytics/ultralytics, 0000-0001-5950- 6979, 0000-0003-3783-7069, AGPL-3.0

work page 2024
[13]

Attention Is All You Need

Attention Is All You Need, Ashish Vaswani and Noam Shazeer and Niki Parmar and Jakob Uszkoreit and Llion Jones and Aidan N. Gomez and Lukasz Kaiser and Illia Polosukhin, 2023, 1706.03762, arXiv, cs.CL, https://arxiv.org/abs/1706.03762,

work page internal anchor Pith review Pith/arXiv arXiv 2023
[14]

org/abs/2107.03006

Structured Denoising Diffusion Models in Discrete State-Spaces, Ja- cob Austin and Daniel D. Johnson and Jonathan Ho and Daniel Tarlow and Rianne van den Berg, 2023, 2107.03006, arXiv, cs.LG, https://arxiv.org/abs/2107.03006,

work page arXiv 2023
[15]

Zhang, T. Y . and Suen, C. Y ., A fast parallel algorithm for thinning digital patterns, 1984, issue date = March 1984, publisher = Association for Computing Machinery, address = New York, NY , USA, 27, 3, 0001- 0782, https://doi.org/10.1145/357994.358023, 10.1145/357994.358023, Commun. ACM, mar, 236–239, num4, parallel algorithm, skeletoniza- tion, thinn...

work page doi:10.1145/357994.358023 1984
[16]

Generative Adversarial Networks, Ian J. Goodfellow and Jean Pouget- Abadie and Mehdi Mirza and Bing Xu and David Warde-Farley and Sherjil Ozair and Aaron Courville and Yoshua Bengio, 2014, 1406.2661, arXiv, stat.ML, https://arxiv.org/abs/1406.2661,

work page internal anchor Pith review Pith/arXiv arXiv 2014
[17]

The HAM10000 dataset, a large collection of multi-source der- matoscopic images of common pigmented skin lesions, 5, 2052- 4463, http://dx.doi.org/10.1038/sdata.2018.161, 10.1038/sdata.2018.161, 1, Scientific Data, publisher=Springer Science and Business Media LLC, Tschandl, Philipp and Rosendahl, Cliff and Kittler, Harald, 2018, month=aug

work page doi:10.1038/sdata.2018.161 2052
[18]

KV ASIR: A Multi-Class Image Dataset for Computer Aided Gastroin- testinal Disease Detection, Pogorelov, Konstantin and Randel, Kristin Ranheim and Griwodz, Carsten and Eskeland, Sigrun Losada and de Lange, Thomas and Johansen, Dag and Spampinato, Concetto and Dang-Nguyen, Duc-Tien and Lux, Mathias and Schmidt, Peter Thelin and Riegler, Michael and Halvor...

work page doi:10.1145/3083187.3083212 2017
[19]

Denoising Diffusion Implicit Models, Jiaming Song and Chen- lin Meng and Stefano Ermon, 2022, 2010.02502, arXiv, cs.LG, https://arxiv.org/abs/2010.02502,

work page internal anchor Pith review Pith/arXiv arXiv 2022
[20]

Denoising Diffusion Probabilistic Models, Jonathan Ho and Ajay Jain and Pieter Abbeel, 2020, 2006.11239, arXiv, cs.LG, https://arxiv.org/abs/2006.11239,

work page internal anchor Pith review Pith/arXiv arXiv 2020
[21]

Deep Residual Learning for Image Recognition, Kaiming He and Xiangyu Zhang and Shaoqing Ren and Jian Sun, 2015, 1512.03385, arXiv, cs.CV, https://arxiv.org/abs/1512.03385,

work page internal anchor Pith review Pith/arXiv arXiv 2015
[22]

Polyak, B. T. and Juditsky, A. B., Acceleration of Stochastic Approxi- mation by Averaging, SIAM Journal on Control and Optimization, 30, 4, 838-855, 1992, 10.1137/0330046, https://doi.org/10.1137/0330046 ,

work page doi:10.1137/0330046 1992
[23]

Decoupled Weight Decay Regularization, Ilya Loshchilov and Frank Hutter, 2019, 1711.05101, arXiv, cs.LG, https://arxiv.org/abs/1711.05101,

work page internal anchor Pith review Pith/arXiv arXiv 2019
[24]

Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation, Liang-Chieh Chen and Yukun Zhu and George Papandreou and Florian Schroff and Hartwig Adam, 2018, 1802.02611, arXiv, cs.CV, https://arxiv.org/abs/1802.02611,

work page internal anchor Pith review Pith/arXiv arXiv 2018
[25]

RefineMask: Towards High-Quality Instance Segmentation with Fine- Grained Features, Gang Zhang and Xin Lu and Jingru Tan and Jianmin Li and Zhaoxiang Zhang and Quanquan Li and Xiaolin Hu, 2021, 2104.08569, arXiv, cs.CV, https://arxiv.org/abs/2104.08569,

work page arXiv 2021
[26]

SegFix: Model-Agnostic Boundary Refinement for Segmentation, Yuhui Yuan and Jingyi Xie and Xilin Chen and Jingdong Wang, 2020, 2007.04269, arXiv, cs.CV, https://arxiv.org/abs/2007.04269,

work page arXiv 2020
[27]

Cootes and C.J

Active Shape Models-Their Training and Application, Computer Vision and Image Understanding, 61, 1, 38- 59, 1995, 1077-3142, https://doi.org/10.1006/cviu.1995.1004, https://www.sciencedirect.com/science/article/pii/S1077314285710041, T.F. Cootes and C.J. Taylor and D.H. Cooper and J. Graham,

work page doi:10.1006/cviu.1995.1004 1995
[28]

Fast Edge Detection Using Structured Forests

Fast Edge Detection Using Structured Forests, Piotr Doll ´ar and C. Lawrence Zitnick, 2014, 1406.5549, arXiv, cs.CV, https://arxiv.org/abs/1406.5549,

work page internal anchor Pith review Pith/arXiv arXiv 2014
[29]

Canny, John, IEEE Transactions on Pattern Analysis and Ma- chine Intelligence, A Computational Approach to Edge Detection, 1986, PAMI-8, 6, 679-698, Image edge detection;Detectors;Machine vision;Shape measurement;Performance analysis;Uncertainty;Gaussian approximation;Signal to noise ratio;Signal synthesis;Feature extrac- tion;Edge detection;feature extra...

work page doi:10.1109/tpami.1986.4767851 1986
[30]

article, Sobel, Irwin and Feldman, Gary, 1973, 01, 271-272, A 3×3 isotropic gradient operator for image processing, Pattern Classification and Scene Analysis

work page 1973
[31]

Topological structural analysis of digitized binary images by border following, Computer Vision, Graphics, and Image Processing, 30, 1, 32- 46, 1985, 0734-189X, https://doi.org/10.1016/0734-189X(85)90016-7, https://www.sciencedirect.com/science/article/pii/0734189X85900167, Satoshi Suzuki and KeiichiA be,

work page doi:10.1016/0734-189x(85)90016-7 1985
[32]

FCN-Transformer Feature Fusion for Polyp Segmentation, ISBN=9783031120534, 1611-3349, http://dx.doi.org/10.1007/978- 3-031-12053-4 65, 10.1007/978-3-031-12053-4 65, bookMedical Image Understanding and Analysis, publisher=Springer International Publishing, Sanderson, Edward and Matuszewski, Bogdan J., 2022, 892–907

work page doi:10.1007/978- 2022
[33]

and Abolmaesumi, Purang and Stoyanov, Danail and Mateus, Diana and Zuluaga, Maria A

”Fan, Deng-Ping and Ji, Ge-Peng and Zhou, Tao and Chen, Geng and Fu, Huazhu and Shen, Jianbing and Shao, Ling”, editor=”Martel, Anne L. and Abolmaesumi, Purang and Stoyanov, Danail and Mateus, Diana and Zuluaga, Maria A. and Zhou, S. Kevin and Racoceanu, Daniel and Joskowicz, Leo”, ”PraNet: Parallel Reverse Attention Network for Polyp Segmentation”, book”...

work page 2020
[34]

HarDNet-DFUS: An Enhanced Harmonically-Connected Network for Diabetic Foot Ulcer Image Segmentation and Colonoscopy Polyp Seg- mentation, Ting-Yu Liao and Ching-Hui Yang and Yu-Wen Lo and Kuan- Ying Lai and Po-Huai Shen and Youn-Long Lin, 2022, 2209.07313, arXiv, eess.IV, https://arxiv.org/abs/2209.07313,

work page arXiv 2022
[35]

He, Jianzhong and Zhang, Shiliang and Yang, Ming and Shan, Yanhu and Huang, Tiejun, IEEE Transactions on Pattern Analy- sis and Machine Intelligence, BDCN: Bi-Directional Cascade Net- work for Perceptual Edge Detection, 2022, 44, 1, 100-113, Im- age edge detection;Task analysis;Bidirectional control;Fuses;Image seg- mentation;Feature extraction;Convolutio...

work page doi:10.1109/tpami.2020.3007074 2022
[36]

U-Net: Convolutional Networks for Biomedical Image Segmentation, Olaf Ronneberger and Philipp Fischer and Thomas Brox, 2015, 1505.04597, arXiv, cs.CV, https://arxiv.org/abs/1505.04597

work page internal anchor Pith review Pith/arXiv arXiv 2015
[37]

Holistically-Nested Edge Detection, Saining Xie and Zhuowen Tu, 2015, 1504.06375, arXiv, cs.CV, https://arxiv.org/abs/1504.06375,

work page internal anchor Pith review Pith/arXiv arXiv 2015
[38]

MaskDiffusion: Exploiting Pre-trained Diffusion Models for Seman- tic Segmentation, Yasufumi Kawano and Yoshimitsu Aoki, 2024, 2403.11194, arXiv, cs.CV, https://arxiv.org/abs/2403.11194,

work page arXiv 2024
[39]

A Deeply Supervised Semantic Segmentation Method Based on GAN, Wei Zhao and Qiyu Wei and Zeng Zeng, 2023, 2310.04081, arXiv, cs.CV, https://arxiv.org/abs/2310.04081,

work page arXiv 2023