Pith. sign in

REVIEW 4 major objections 3 minor 34 references

Deep Learning-Based Desikan-Killiany Parcellation of the Brain Using Diffusion MRI

T0 review · 4 major / 3 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A two-stage deep network can parcellate the cerebral cortex into Desikan-Killiany regions directly from diffusion MRI, without an anatomical scan or inter-modality registration.

desk verdict Plausible dMRI-only DK parcellation idea, but the supplied full text is a different paper and the abstract's claims are unverified; the registration-free premise depends on how the ground-truth labels were made. read the letter →

arxiv 2508.07815 v1 pith:C55D2NPA submitted 2025-08-11 eess.IV

Yousef Sadegheih , Dorit Merhof This is my paper
classification eess.IV
keywords brainparcellationDesikan-KillianyatlasdiffusionMRIdeeplearningsegmentationregistration-freehierarchicalnetworktensorimaging
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 claims that cortical parcellation according to the Desikan-Killiany atlas can be performed directly in diffusion MRI (dMRI) space using a deep learning network, removing the need for a T1-weighted anatomical scan and the registration between modalities. The framework is a two-stage hierarchical segmentation network: a coarse stage labels broad brain regions, then a refinement stage delineates the full set of Desikan-Killiany subregions. The network takes four diffusion-derived parameter maps as input—fractional anisotropy, trace, sphericity, and maximum eigenvalue—chosen through an ablation study. On the Human Connectome Project and Consortium for Neuropsychiatric Phenomics datasets, the authors report higher Dice similarity coefficients than existing dMRI-based methods, with robust generalization across resolutions and acquisition protocols and more homogeneous parcels. If correct, this simplifies neuroimaging workflows and argues that dMRI alone carries sufficient information for standard cortical parcellation.

What carries the argument

The central mechanism is a hierarchical two-stage segmentation network: the first stage produces a coarse volumetric segmentation of large brain regions, and the second stage, conditioned on that coarse output, refines each region into the finer Desikan-Killiany subregions. The input is a fixed set of four scalar dMRI maps—fractional anisotropy, trace, sphericity, and maximum eigenvalue—selected by an ablation study as the combination maximizing Dice. The network is trained and evaluated on reference DK labels expressed in diffusion space, using Dice similarity and intra-region homogeneity (relative standard deviation) as metrics.

What would settle it

Compare the proposed dMRI-only parcellation against reference labels that were created manually in dMRI space, without any T1-to-dMRI warping, for a small set of subjects; if the Dice advantage over T1-warped baselines disappears or reverses, the registration-free advantage is an artifact of the reference labels. Alternatively, on subjects with deliberately misaligned T1 and dMRI data, check whether the method's Dice advantage over T1-based pipelines grows with the misalignment; if it does not, the 'avoids registration error' mechanism is not supported.

Watch

Extended reading notes

Core claim

The authors claim that a single end-to-end trained two-stage convolutional network, operating on four scalar diffusion-tensor-derived maps, produces Desikan-Killiany parcellations in native dMRI space with accuracy meeting or exceeding that of pipelines that segment a T1-weighted image and warp the labels into diffusion space. The first stage outputs a coarse parcellation into broad brain regions; the second stage refines each coarse region into the fine Desikan-Killiany subregions. The input combination of fractional anisotropy, trace, sphericity, and maximum eigenvalue was selected by exhaustive ablation as the most accurate. The evaluation on HCP and CNP datasets reports superior Dice sim

Load-bearing premise

The reference Desikan-Killiany labels in diffusion space are an accurate training and evaluation target; if they were created by warping T1-based parcellations, the registration error the method claims to avoid is baked into both the training signal and the Dice metric.

Editorial extensions

If this is right

  • If the claim holds, neuroimaging pipelines can obtain standard cortical parcellations from a dMRI acquisition alone, saving scan time and avoiding T1-dMRI registration artifacts.
  • Diffusion-only retrospective datasets, where no anatomical image was collected, could be labeled for large-scale studies of connectivity and microstructure.
  • The identified optimal map combination suggests which scalar dMRI contrasts carry the most parcellation-relevant information, guiding future feature selection in dMRI segmentation.
  • The coarse-to-fine hierarchical design may transfer to other cortical atlases or to subcortical structures, reducing the need for bespoke architectures.
  • The reported robustness across resolutions and protocols implies the network could be applied across scanners without resampling to a common grid, lowering preprocessing burden.

Reading between the lines

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

  • Editorial note: the full-text body supplied is a different paper on zero-shot anomaly detection; the pith above is extracted from the title and abstract, which describe the DK parcellation framework.
  • The main unresolved risk is the provenance of the reference DK labels: if they were generated by warping T1-based parcellations into dMRI space, the training target contains the very registration error the method claims to avoid, making the reported 'registration-free' advantage partially circular.
  • The four-map combination (FA, trace, sphericity, maximum eigenvalue) is likely dataset-dependent; on acquisitions with fewer gradient directions or lower b-values, some of these maps become noisy and the optimal set may shift—a testable extension of the ablation.
  • If the method proves robust on independent diffusion-only datasets with manually defined DK ground truth, it would strengthen the case that cortical parcellation can be derived from microstructural contrasts alone, potentially enabling post-mortem or fetal studies where T1 contrast is poor.
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 / 3 minor

Summary. The submission, identified by its title and abstract as arXiv:2508.07815, claims a deep-learning framework for Desikan-Killiany (DK) parcellation directly from diffusion MRI (dMRI) data. The abstract asserts a hierarchical two-stage network, an ablation study identifying an optimal combination of fractional anisotropy, trace, sphericity, and maximum eigenvalue, superior Dice Similarity Coefficients compared with state-of-the-art models on HCP and CNP datasets, and robust generalization across resolutions and acquisition protocols. However, the supplied full text is an entirely different paper on zero-shot anomaly detection (ACD-CLIP), with no dMRI content, no DK parcellation methods, no experimental results, and no description of how reference labels were generated. The abstract's central empirical claims are therefore unverifiable from the submitted manuscript.

Significance. If the claimed method worked, direct DK parcellation from dMRI alone would be a useful contribution to neuroimaging: it could reduce reliance on T1 acquisition and inter-modality registration, and the DK atlas provides an external, well-established benchmark rather than a self-defined target. The two-stage coarse-to-fine architecture and the ablation over diffusion-derived maps are also reasonable design elements. The manuscript offers no verifiable evidence for these claims, however: there are no quantitative results, no methods description, no label-generation protocol, and no statistical comparisons in the supplied text. The intended contribution is potentially significant, but as submitted the paper is not scientifically assessable.

major comments (4)
  1. [Title/Abstract vs. Full Text] The manuscript body is not the paper announced by the title and abstract. The full text is an unrelated paper titled 'ACD-CLIP: Decoupling Representation and Dynamic Fusion for Zero-Shot Anomaly Detection' with its own abstract, methodology, experiments on MVTec-AD and medical anomaly benchmarks, and references. There is no description of the DK parcellation network, no dMRI experiments, no HCP/CNP results, and no discussion of registration or parcellation. This is not a local presentation defect: the entirety of the methods and results supporting the abstract's claims is missing. The central claim cannot be checked in any way.
  2. [Abstract, 'superior Dice Similarity Coefficients'] The abstract asserts superiority over existing state-of-the-art models but reports no Dice values, no confidence intervals, no standard deviations, and no statistical tests. It also does not name the baselines or define the evaluation protocol (cortical regions, volumetric vs. surface, overlap measures). In the absence of any tables or results in the submitted text, this is an unsupported assertion rather than an empirical finding. A revised manuscript must provide the full quantitative comparison with per-region and aggregate metrics.
  3. [Abstract, reference-label generation / registration-free claim] The central selling point is a 'registration-free' pipeline that uses only dMRI data. The abstract does not explain how ground-truth DK labels in dMRI space were obtained. The standard approach is to parcellate a T1-weighted image and warp the labels into diffusion space; if this was used, the training and evaluation targets are produced by exactly the inter-modality registration the paper claims to avoid. Registration errors in the reference labels would contaminate the training signal and the Dice metric, making the 'registration-free' advantage at least partially circular. The full text must describe label-generation steps and quantify their accuracy; currently this information is absent.
  4. [Abstract, ablation study and input-map selection] The abstract states that an extensive ablation study identified FA, trace, sphericity, and maximal eigenvalue as the optimal combination, and that this improves 'parellation accuracy.' No ablation table, protocol, or definition of 'optimal' is provided in the supplied text. Because the choice of input maps is a free parameter of the method, the claimed optimality cannot be assessed, and the omission prevents reproduction of the framework. Detailed ablation results are required.
minor comments (3)
  1. [Abstract, typo] 'parellation' should be 'parcellation' in 'enhances parellation accuracy.'
  2. [Full-text references] The reference list in the supplied body belongs to the ACD-CLIP anomaly-detection paper and contains no citations to diffusion MRI, Desikan-Killiany parcellation, or neuroimaging registration. If the correct manuscript is resubmitted, the bibliography must be replaced accordingly.
  3. [Code availability] The abstract says the implementation is publicly available at github.com/xmindflow/DKParcellationdMRI, but the manuscript gives no documentation, usage instructions, or version information. The link cannot be validated from the submission.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity is demonstrable from the available abstract; the supplied full text is a different paper, so no derivation chain can be inspected.

full rationale

The submitted manuscript consists of an abstract for arXiv:2508.07815 (dMRI-based DK parcellation) followed by an unrelated full text for a different paper (ACD-CLIP, zero-shot anomaly detection). No methods, equations, or experimental details for the dMRI parcellation paper are present. Under the hard rule that circularity must be exhibited by quoting the paper's own equations or explicit reductions, no specific circular step can be identified from the abstract alone. The abstract's claim of a 'registration-free' pipeline and superior Dice does not, by itself, define the target in terms of the model's outputs or fit any parameter that is then called a prediction. The concern raised about ground-truth label generation (whether reference DK labels were produced via T1-based warping into dMRI space) is a validity/verifiability issue, not a demonstrated circularity: it would only be circular if the paper defined its evaluation metric in terms of those labels in a way that forces the result, which cannot be shown from the available text. The full text's content is irrelevant to the abstract's paper, reinforcing that no derivation chain is available for analysis. Therefore, the correct finding is no significant circularity (score 0).

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the quality of the dMRI-space reference labels (external atlas plus a transform), the information content of the four selected maps, and standard supervised-learning assumptions. The abstract supplies none of the corresponding validation evidence; the body that could carry it was not available.

free parameters (1)
  • Input parameter-map combination (FA, trace, sphericity, maximum eigenvalue) = FA + trace + sphericity + max eigenvalue
    The abstract says an extensive ablation study 'identified' this optimal combination; selection on the evaluation datasets would make the four-map input a data-fitted choice rather than an independent design.
assumptions (3)
  • domain assumption Reference DK labels in dMRI space are valid and accurate for every HCP/CNP subject
    The network is trained and evaluated against DK parcellations in diffusion space; the abstract does not state how these labels were generated. Any registration or atlas-transfer used to build them injects its own error into both training and evaluation.
  • domain assumption The four dMRI-derived maps carry enough information to delineate dozens of cortical regions
    The whole approach rests on FA, trace, sphericity, and maximum eigenvalue sufficing for region discrimination; no information-theoretic or anatomical argument appears in the abstract.
  • domain assumption No leakage between training and test sets across HCP and CNP
    Unverifiable from the abstract; subject overlap between the two datasets would inflate reported Dice scores.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Learning-Based Desikan-Killiany Parcellation of the Brain Using Diffusion MRI." pith.science (2026). https://pith.science/paper/C55D2NPA

@misc{pith2026250807815,
  author       = {Pith},
  title        = {Pith review of: Deep Learning-Based Desikan-Killiany Parcellation of the Brain Using Diffusion MRI},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C55D2NPA}},
  note         = {Machine review of arXiv:2508.07815}
}
read the original abstract

Accurate brain parcellation in diffusion MRI (dMRI) space is essential for advanced neuroimaging analyses. However, most existing approaches rely on anatomical MRI for segmentation and inter-modality registration, a process that can introduce errors and limit the versatility of the technique. In this study, we present a novel deep learning-based framework for direct parcellation based on the Desikan-Killiany (DK) atlas using only diffusion MRI data. Our method utilizes a hierarchical, two-stage segmentation network: the first stage performs coarse parcellation into broad brain regions, and the second stage refines the segmentation to delineate more detailed subregions within each coarse category. We conduct an extensive ablation study to evaluate various diffusion-derived parameter maps, identifying an optimal combination of fractional anisotropy, trace, sphericity, and maximum eigenvalue that enhances parellation accuracy. When evaluated on the Human Connectome Project and Consortium for Neuropsychiatric Phenomics datasets, our approach achieves superior Dice Similarity Coefficients compared to existing state-of-the-art models. Additionally, our method demonstrates robust generalization across different image resolutions and acquisition protocols, producing more homogeneous parcellations as measured by the relative standard deviation within regions. This work represents a significant advancement in dMRI-based brain segmentation, providing a precise, reliable, and registration-free solution that is critical for improved structural connectivity and microstructural analyses in both research and clinical applications. The implementation of our method is publicly available on github.com/xmindflow/DKParcellationdMRI.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 26 canonical work pages

  1. [1]

    ACD-CLIP: Decoupling Representation and Dynamic Fusion for Zero-Shot Anomaly Detection

    INTRODUCTION Zero-Shot Anomaly Detection (ZSAD) adapts Vision-Language Models (VLMs) [1, 2] like CLIP [3] to circumvent the ex- tensive training data required by traditional methods [4, 5]. The dominant paradigm uses text prompts, evolving from hand-crafted ensembles to learnable prompts tuned on aux- iliary data (WinCLIP [6]). These advanced methods have...

    work page Pith review arXiv 2026

  2. [2]

    METHODOLOGY As illustrated in Figure 2, our ACD-CLIP framework intro- duces two core innovations to adapt VLMs for ZSAD:(1)a Conv-LoRA Adapterto instill local inductive biases for fine- grained representation, and(2)aDynamic Fusion Gateway for adaptive cross-modal fusion. 2.1. Hierarchical Feature Adaptation with Local Priors To mitigate feature entanglem...

    work page

  3. [3]

    EXPERIMENTS To demonstrate the efficacy and robustness of our ACD-CLIP framework, we conducted a comprehensive evaluation. This involved benchmarking our model against state-of-the-art (SOTA) competitors across a wide range of datasets and per- forming in-depth ablation studies to quantify the contribution of each individual component. 3.1. Experimental S...

    work page

  4. [4]

    + Conv-LoRA Adapter 89.1 (+6.8) 84.1 (+2.9)

    work page

  5. [5]

    + DFG 87.6 (+5.3) 85.9 (+4.7)

    work page

  6. [6]

    This superiority in capturing precise anomaly boundaries is a direct result of our architectural co-design

    Ours (ACD-CLIP) 91.7 90.9 AnomalyCLIP. This superiority in capturing precise anomaly boundaries is a direct result of our architectural co-design. Trained only on industrial data, ACD-CLIP shows strong cross-domain generalization, achieving 90.4% and 96.6% pixel-level AUROC on ClinicDB and BrainMRI, respectively. Qualitative results (Figure 3) substantiat...

    work page

  7. [7]

    CONCLUSION We proposed ACD-CLIP, an Architectural Co-Design frame- work addressing VLM limitations in Zero-Shot Anomaly De- tection. By synergizing a parameter-efficientConv-LoRA adapterfor local priors with aDynamic Fusion Gateway for adaptive cross-modal interaction, our model extracts fine- grained, context-aware features. ACD-CLIP achieves state- of-t...

    work page

  8. [8]

    Flamingo: a vi- sual language model for few-shot learning,

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al., “Flamingo: a vi- sual language model for few-shot learning,”Advances in neu- ral information processing systems, vol. 35, pp. 23716–23736, 2022

    work page 2022

Show all 34 references
  1. [9]

    Visual instruction tuning,

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee, “Visual instruction tuning,”Advances in neural information processing systems, vol. 36, pp. 34892–34916, 2023

    2023

  2. [10]

    Learning trans- ferable visual models from natural language supervision,

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al., “Learning trans- ferable visual models from natural language supervision,” in International conference on machine learnin...

    2021

  3. [11]

    Towards total recall in industrial anomaly detection,

    Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Sch¨olkopf, Thomas Brox, and Peter Gehler, “Towards total recall in industrial anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, 2022, pp. 14318–14328

    2022

  4. [12]

    Padim: a patch distribution modeling frame- work for anomaly detection and localization,

    Thomas Defard, Aleksandr Setkov, Angelique Loesch, and Ro- maric Audigier, “Padim: a patch distribution modeling frame- work for anomaly detection and localization,” inInternational conference on pattern recognition. Springer, 2021, pp. 475– 489

    2021

  5. [13]

    Winclip: Zero- /few-shot anomaly classification and segmentation,

    Jongheon Jeong, Yang Zou, Taewan Kim, Dongqing Zhang, Avinash Ravichandran, and Onkar Dabeer, “Winclip: Zero- /few-shot anomaly classification and segmentation,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 19606–19616

    2023

  6. [14]

    Anomalyclip: Object-agnostic prompt learn- ing for zero-shot anomaly detection,

    Qihang Zhou, Guansong Pang, Yu Tian, Shibo He, and Jim- ing Chen, “Anomalyclip: Object-agnostic prompt learn- ing for zero-shot anomaly detection,”arXiv preprint arXiv:2310.18961, 2023

    2023

  7. [15]

    Adaclip: Adapt- ing clip with hybrid learnable prompts for zero-shot anomaly detection,

    Yunkang Cao, Jiangning Zhang, Luca Frittoli, Yuqi Cheng, Weiming Shen, and Giacomo Boracchi, “Adaclip: Adapt- ing clip with hybrid learnable prompts for zero-shot anomaly detection,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 55–72

    2024

  8. [16]

    Clip-ad: A language-guided staged dual-path model for zero- shot anomaly detection,

    Xuhai Chen, Jiangning Zhang, Guanzhong Tian, Haoyang He, Wuhao Zhang, Yabiao Wang, Chengjie Wang, and Yong Liu, “Clip-ad: A language-guided staged dual-path model for zero- shot anomaly detection,” inInternational Joint Conference on Artificial Intelligence. Springer, 2024, pp. 17–33

    2024

  9. [17]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa De- hghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al., “An image is worth 16x16 words: Transformers for image recognition at scale,”arXiv preprint a...

    2010 arXiv

  10. [18]

    Fully convolutional networks for semantic segmentation,

    Jonathan Long, Evan Shelhamer, and Trevor Darrell, “Fully convolutional networks for semantic segmentation,” inPro- ceedings of the IEEE conference on computer vision and pat- tern recognition, 2015, pp. 3431–3440

    2015

  11. [19]

    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. Springer, 2015, pp. 234–241

    2015

  12. [20]

    Parameter-efficient transfer learning for nlp,

    Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly, “Parameter-efficient transfer learning for nlp,” inInternational conference on machine learning. PMLR, 2019, pp. 2790–2799

    2019

  13. [21]

    Lora: Low-rank adaptation of large language models.,

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al., “Lora: Low-rank adaptation of large language models.,”ICLR, vol. 1, no. 2, pp. 3, 2022

    2022

  14. [22]

    Convolutional bypasses are better vision transformer adapters,

    Shibo Jie, Zhi-Hong Deng, Shixuan Chen, and Zhijuan Jin, “Convolutional bypasses are better vision transformer adapters,” inECAI 2024, pp. 202–209. IOS Press, 2024

    2024

  15. [23]

    Convolution meets lora: Parameter efficient finetuning for segment anything model,

    Zihan Zhong, Zhiqiang Tang, Tong He, Haoyang Fang, and Chun Yuan, “Convolution meets lora: Parameter efficient finetuning for segment anything model,”arXiv preprint arXiv:2401.17868, 2024

    2024 arXiv

  16. [24]

    Focal loss for dense object detection,

    Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Doll´ar, “Focal loss for dense object detection,” inPro- ceedings of the IEEE international conference on computer vi- sion, 2017, pp. 2980–2988

    2017

  17. [25]

    V-net: Fully convolutional neural networks for volumetric medical image segmentation,

    Fausto Milletari, Nassir Navab, and Seyed-Ahmad Ahmadi, “V-net: Fully convolutional neural networks for volumetric medical image segmentation,” in2016 fourth international conference on 3D vision (3DV). Ieee, 2016, pp. 565–571

    2016

  18. [26]

    Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection,

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger, “Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection,” inProceedings of the IEEE/CVF conference on computer vision and pattern recog- nition, 2019, pp. 9592–9600

    2019

  19. [27]

    Spot-the-difference self-supervised pre- training for anomaly detection and segmentation,

    Yang Zou, Jongheon Jeong, Latha Pemula, Dongqing Zhang, and Onkar Dabeer, “Spot-the-difference self-supervised pre- training for anomaly detection and segmentation,” inEuropean conference on computer vision. Springer, 2022, pp. 392–408

    2022

  20. [28]

    Vt-adl: A vision transformer network for image anomaly detection and localization,

    Pankaj Mishra, Riccardo Verk, Daniele Fornasier, Claudio Pi- ciarelli, and Gian Luca Foresti, “Vt-adl: A vision transformer network for image anomaly detection and localization,” in 2021 IEEE 30th International Symposium on Industrial Elec- tronics (ISIE). IEEE, 2021, pp. 01–06

    2021

  21. [29]

    Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions,

    Stepan Jezek, Martin Jonak, Radim Burget, Pavel Dvorak, and Milos Skotak, “Deep learning-based defect detection of metal parts: evaluating current methods in complex conditions,” in 2021 13th International congress on ultra modern telecommu- nications and control systems and w...

    2021

  22. [30]

    A coarse-to-fine model for rail surface defect detection,

    Haomin Yu, Qingyong Li, Yunqiang Tan, Jinrui Gan, Jianzhu Wang, Yangli-ao Geng, and Lei Jia, “A coarse-to-fine model for rail surface defect detection,”IEEE Transactions on In- strumentation and Measurement, vol. 68, no. 3, pp. 656–666, 2018

    2018

  23. [31]

    Bmad: Benchmarks for medical anomaly detection,

    Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxi- ang Zhang, and Xingyu Li, “Bmad: Benchmarks for medical anomaly detection,” inProceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition, 2024, pp. 4042–4053

    2024

  24. [32]

    Automated polyp detection in colonoscopy videos using shape and context information,

    Nima Tajbakhsh, Suryakanth R Gurudu, and Jianming Liang, “Automated polyp detection in colonoscopy videos using shape and context information,”IEEE transactions on medical imag- ing, vol. 35, no. 2, pp. 630–644, 2015

    2015

  25. [33]

    Wm- dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians,

    Jorge Bernal, F Javier S ´anchez, Gloria Fern ´andez-Esparrach, Debora Gil, Cristina Rodr´ıguez, and Fernando Vilari˜no, “Wm- dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians,”Computerized medical imaging and graphics, vol....

    2015

  26. [34]

    Kvasir-seg: A segmented polyp dataset,

    Debesh Jha, Pia H Smedsrud, Michael A Riegler, P ˚al Halvorsen, Thomas De Lange, Dag Johansen, and H ˚avard D Johansen, “Kvasir-seg: A segmented polyp dataset,” inInter- national conference on multimedia modeling. Springer, 2019, pp. 451–462

    2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.