Pith. sign in

REVIEW 4 major objections 4 minor 56 references

CFR-Net:Collaborative Feature Refinement Network for Medical Image Anomaly Detection

T0 review · 4 major / 4 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 A shared-refinement teacher-student network lifts medical anomaly localization to a mean AUROC of 98.61 percent.

desk verdict A plausible reverse-distillation variant for medical anomaly detection whose headline SOTA claim rests on a benchmark table assembled from incommensurable sources. read the letter →

arxiv 2607.11509 v2 pith:4N5OQ5G4 submitted 2026-07-13 cs.CV

Zihan Nie , Muhao Xu , Wei Feng , Yuan Cui , Hua Wei , Sijie Niu , Yi Wan , Xunbin Wei
show 2 more authors
Weiye Song Zongyuan Ge
This is my paper
classification cs.CV
keywords anomalydetectionmedicalimagingteacher-studentnetworkfeaturerefinementknowledgedistillationunsupervisedlearninglocalization
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

This paper proposes CFR-Net, a teacher-student framework for detecting anomalies in medical images using only normal data for training. The key claim is that jointly refining a frozen ImageNet-pretrained teacher and a trainable student with a shared multi-path module before decoding, then enforcing cross-space consistency after decoding, adapts generic visual features to the medical domain and yields state-of-the-art anomaly classification and localization. On six medical imaging benchmarks, the method reports the best results in seven of nine tasks, with a mean localization AUROC of 98.61% and a mean classification AUROC of 87.23%. The design is particularly strong at localizing spatially precise anomalies such as tumors and retinal layer distortions.

What carries the argument

The load-bearing component is the Multi-Path Feature Refinement Module (MPFRM), a shared-parameter operator applied to both teacher and student encoder features. MPFRM combines three branches: a variance-weighted branch that highlights patch-level feature dispersion, a multi-scale pyramid branch that aggregates context across receptive field sizes, and an asymmetric convolution branch that captures orientation-sensitive structural patterns. These branches are fused via residual connections. Crucially, the same MPFRM is applied to both streams, imposing common refinement rules, and the decoded features are then trained with a crossed teacher-student consistency loss.

What would settle it

Run CFR-Net and all compared anomaly detection methods on the same six medical datasets using identical preprocessing, data splits, and evaluation code; if the mean localization AUROC gap between CFR-Net and the best baseline narrows to below statistical significance, the central claim is not established.

Watch

Extended reading notes

Core claim

CFR-Net's central discovery is that applying the same refinement operator to both the frozen teacher's generic features and the student's medical-domain-adapted features before decoding—then comparing each decoded stream with the complementary encoder feature space—substantially improves unsupervised anomaly detection in medical images. The reported gains are not incremental: adding the shared Multi-Path Feature Refinement Module (MPFRM) improves localization AUROC by over 7 percentage points over the teacher-student baseline, and the full framework reaches a mean localization AUROC of 98.61% across five localization tasks. The paper argues that this collaborative refinement mitigates the do

Load-bearing premise

The claim of state-of-the-art performance depends on comparing CFR-Net's numbers to baseline scores taken from other papers or an external benchmark, rather than re-running all methods under one controlled evaluation protocol.

Editorial extensions

If this is right

  • If the reported results hold, normal-only trained medical anomaly detectors can achieve near-clinical localization accuracy, potentially reducing the need for labeled abnormal data in screening applications.
  • The shared-refinement design suggests a general recipe for adapting ImageNet-pretrained features to other specialized imaging domains, beyond medical data.
  • The variance-sensitive weighting and dynamic 'homework set' reorganization point to a stable training recipe for teacher-student distillation under one-class settings.
  • The improvement in localization is consistently larger than in classification, which may guide future methods to prioritize spatially precise feature refinement.
  • The framework is robust across several encoder backbones, implying it can be deployed with different pretrained networks without re-engineering.

Reading between the lines

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

  • Because the paper does not release code, the practical impact depends on independent reimplementation; a unified evaluation harness with identical splits would verify whether the headline numbers generalize beyond the reported runs.
  • The shared MPFRM implicitly imposes a common inductive bias on both streams; a testable prediction is that using separate refinement parameters for teacher and student would degrade performance, which would confirm the value of the shared constraint.
  • The dynamic 'homework set' reorganization resembles a curriculum over data subsets; it likely stabilizes training by reducing overfitting to a fixed random subset, a mechanism that could transfer to other one-class learning tasks.
  • Extending the architecture to 3D volumes, as the authors suggest, may require rethinking the asymmetric branch, since orientation sensitivity in 3D becomes more complex; this is a natural next test for the framework's generality.
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 / 4 minor

Summary. The paper proposes CFR-Net, a teacher-student framework for unsupervised medical image anomaly detection. A frozen ImageNet-pretrained Wide-ResNet50-2 teacher encoder and a trainable student encoder extract hierarchical features; both streams are processed before decoding by a shared Multi-Path Feature Refinement Module (MPFRM) with variance-weighted, multi-scale pyramid, and asymmetric convolution branches. After decoding, the paper introduces a crossed teacher-student consistency loss with variance-sensitive feature-level weighting and a dynamic 'homework set' re-partitioning strategy. Experiments are reported on six medical benchmarks (HIS, OCT17, APTOS, BrainMRI, LiverCT, RESC) using AUROC for classification and localization, claiming best results in seven of nine tasks and a mean localization AUROC of 98.61%. The paper also includes progressive ablations, backbone comparisons, and a sensitivity analysis of the balancing parameter lambda.

Significance. If the empirical claims hold, CFR-Net would be a competitive normal-only medical anomaly detector, with particularly strong localization performance. The core idea—sharing refinement between teacher and student features before decoding and applying crossed consistency after decoding—is clearly motivated and architecturally distinct from standard reverse distillation. The paper covers six modalities, provides component-level ablations, and includes qualitative localization maps. However, the evaluation does not currently establish the headline claim: the table mixes BMAD-copied, literature-copied, and locally run baselines; all numbers are single runs without uncertainty; the test-time anomaly score is not specified; and the paper ships no code. The contribution is plausible, but it needs a unified, reproducible evaluation protocol before the reported state-of-the-art claim can be accepted.

major comments (4)
  1. [§4.2, Table 1] The headline 'best in seven of nine tasks' rests on a table assembled from incompatible protocols. Section 4.2 states that CFLOW-AD, RD4AD, PatchCore, and MKD on Brain MRI, Liver CT, Retinal OCT, and HIS are copied from BMAD, and APTOS baselines are sourced from other papers; only DiAD, msflow, and URD are rerun 'under our experimental setup.' The paper never states that CFR-Net's own runs used BMAD's exact preprocessing, train/test split, and scoring code, nor does it give the anomaly-score formula used at test time. Differences of 1–2 AUROC points (e.g., APTOS AC 96.07 vs. msflow 94.90; RESC AL 98.57 vs. URD 96.74) can easily be artifacts of image size, normalization, or score aggregation. Please provide a unified evaluation in which all baselines are rerun, or at minimum document the exact CFR-Net preprocessing/scoring and confirm it matches the protocol of the cited baseline numbers.
  2. [§5.1, Tables 1–3] No uncertainty quantification is provided. Every number in Tables 1–3 is a single run with no error bars, seeds, or significance tests. Several claimed improvements are small—e.g., OCT17 AC 99.63 vs. PatchCore 98.56, APTOS AC 96.07 vs. msflow 94.90, RESC AL 98.57 vs. URD 96.74. Without repeated runs and an appropriate paired test, the 'best' claims are not statistically supported. Please report mean ± std over at least 3–5 seeds and, where possible, a paired significance test against the strongest baselines.
  3. [§5.2, Fig. 6] Hyperparameters appear to be selected on the evaluation benchmarks. Figure 6 tunes λ on the same datasets used in Table 1, and the α/β schedules (Eq. 13), homework-set fraction (50%), MPFRM patch size, and kernel sizes are fixed without a held-out validation procedure. This creates a risk that the reported numbers overfit the test sets. Please add a validation split or a nested selection protocol, or alternatively show that the main results are stable across a range of hyperparameter choices rather than only the benchmark-selected values.
  4. [§3.3, §4.3] The test-time anomaly scoring mechanism is not defined. Equations (10)–(19) define training losses only. It is unclear how per-pixel anomaly localization maps and image-level classification scores are derived from the decoded features, how the teacher and student streams are combined at inference, and whether BMAD's scoring protocol is used. Without these equations, the AUROC numbers in Table 1 are not reproducible and the comparison with BMAD-copied baselines is not verifiable. Please add the anomaly score formulas and evaluation details.
minor comments (4)
  1. [Figure 2] Typos in the figure: 'Varience' should be 'Variance' and 'Kernal' should be 'Kernel'. Also, the paper inconsistently writes 'CFLOW-AD' in the text and 'CFlowAD' in Table 1.
  2. [§5.2, Figure 5] The claim that the total MPFRM gain exceeds the sum of decomposed gains would be more convincing with a table of absolute AUROC values and error bars; the stacked bar chart alone does not substantiate the 'synergistic' interaction among branches.
  3. [§4.3] The dynamic homework-set re-partitioning (every five epochs, 50% random subset) introduces stochasticity beyond standard training. Please report the random seeds used and the number of runs; otherwise the reported single numbers cannot be interpreted.
  4. [Tables 1–3] The captions should explicitly mark which entries are copied from prior papers, which are from BMAD, and which were rerun in this work, so that readers can distinguish protocol-compatible numbers from literature values.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: CFR-Net's claims are empirical benchmark results; no equation-level reduction or load-bearing self-citation.

full rationale

CFR-Net is an empirical architecture paper rather than a derivational one. The proposed method defines a teacher-student refinement pipeline with explicit losses (cosine discrepancy in Eq. 10, cross-space consistency in Eqs. 11-12, and the weighted objective in Eq. 19), but these are training objectives, not formulas that by construction produce the reported AUROC values in Table 1. The reported numbers come from running the trained model on unseen test images, so no fitted parameter is renamed as a prediction. Hand-set hyperparameters (alpha, beta, lambda, patch size, kernel sizes) are ablated on the evaluation benchmarks, which raises a tuning concern but does not make the comparison true by definition. The paper's self-citations ([37], [52]) appear only in related-work discussion and are not load-bearing for the central claim. The benchmark-comparability issue--some baseline numbers are copied from BMAD or other papers while others are rerun under the authors' setup--is a correctness and evidence concern, not a circularity concern. No equation can be exhibited where a claimed prediction reduces to its input by construction, and no load-bearing argument rests on an unverified self-citation. Thus the appropriate finding is no significant circularity.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

CFR-Net's central performance claim is empirical and rests on a large number of hand-set hyperparameters (α, β, λ, homework fraction, patch/kernel sizes) and on domain assumptions about how anomalies present in medical images and how ImageNet features transfer. No new physical entities are introduced.

free parameters (7)
  • alpha schedule (alpha_initial=0.3, alpha_final=0.2) = 0.3 -> 0.2
    Eq. 13; controls the variance-sensitive feature-level weight exponent. Chosen by hand; no ablation in main text.
  • beta schedule (beta_initial=0.1, beta_final=0.05) = 0.1 -> 0.05
    Eq. 13; threshold below which the cross-space loss is doubled. Chosen by hand.
  • lambda = 0.1
    Eq. 19; balances the weighted and unweighted amplified consistency losses. Selected via sensitivity analysis (Fig. 6) on the same benchmarks.
  • homework set fraction = 0.5
    Sec. 3.3; every five epochs, 50% of the training set is re-sampled as the homework set. Chosen without a reported ablation.
  • MPFRM patch size = 4x4
    Sec. 4.3; patch partition size for the variance-weighted branch.
  • pyramid kernel sizes = 3x3, 5x5, 7x7
    Sec. 4.3; multi-scale pyramid branch kernel sizes.
  • asymmetric kernel sizes = 1x3, 3x1, 3x3
    Sec. 4.3; horizontal/vertical/harmony convolution kernel sizes.
assumptions (4)
  • domain assumption ImageNet-pretrained Wide-ResNet50-2 features provide a useful generic visual basis for normal medical image patterns.
    Invoked in Sec. 1 and 3.1; the method never separately validates that these features align with medical anatomy; the only evidence is final AUROC.
  • domain assumption Anomalies in medical images appear as fine-grained local shifts, multi-scale contextual mismatches, and orientation-sensitive structural deviations, and these three axes are sufficient.
    Stated in the abstract and Sec. 3.2; motivates the three MPFRM branches but is not proven by data.
  • domain assumption Baseline results from BMAD and prior literature are directly comparable to the authors' own runs under identical splits, preprocessing, and metrics.
    Sec. 4.2; several baseline numbers are quoted rather than re-run; if protocols differ, Table 1 comparisons are invalid.
  • domain assumption Cosine discrepancy between flattened feature maps is an adequate distance for cross-space consistency.
    Sec. 3.3, Eq. 10; standard in the field but an assumption about the geometry of the two feature spaces.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CFR-Net:Collaborative Feature Refinement Network for Medical Image Anomaly Detection." pith.science (2026). https://pith.science/paper/4N5OQ5G4

@misc{pith2026260711509,
  author       = {Pith},
  title        = {Pith review of: CFR-Net:Collaborative Feature Refinement Network for Medical Image Anomaly Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4N5OQ5G4}},
  note         = {Machine review of arXiv:2607.11509}
}
read the original abstract

Medical image anomaly detection remains challenging because networks pretrained on natural images often exhibit limited adaptability to medical images, where abnormal patterns appear as fine-grained local shifts, multi-scale contextual mismatches, and orientation-sensitive structural deviations. To address this, we propose the Collaborative Feature Refinement Network (CFR-Net), which combines shared teacher-student feature refinement before decoding with cross-space consistency after decoding. CFR-Net refines frozen teacher features and trainable student features using a Multi-Path Feature Refinement Module (MPFRM) with shared parameters, imposing common multi-path refinement rules on generic visual references and representations adapted to the medical domain, thereby mitigating domain discrepancy while modeling local, multi-scale, and orientation-sensitive feature characteristics. A variance-sensitive objective and dynamic ''homework set'' reorganization further support layer-adaptive consistency learning. Experiments on medical benchmarks show that CFR-Net achieves competitive anomaly classification and strong anomaly localization performance when trained on normal data.

Figures

Figures reproduced from arXiv: 2607.11509 by the authors.

Figure 1
Figure 1. Overall architecture of CFR-Net. A frozen teacher encoder and a trainable stu [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Multi-Path Feature Refinement Module (MPFRM). The module refines teacher [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Qualitative comparison of anomaly localization maps. From left to right: input [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Visualization of anomaly detection maps for classification tasks. The rows show [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Average AUROC gains contributed by different MPFRM components for anomaly [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Impact of hyperparameter λ. The loss balancing parameter λ exerts a nonlinear influence on model performance. 6 Conclusion This study presents CFR-Net, a collaborative feature refinement framework for medical im￾age anomaly detection. CFR-Net refines teacher and studen…

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 5 linked inside Pith

  1. [1]

    Diagnostic accuracy of deep learning in medical imaging: a systematic review and meta-analysis.NPJ digital medicine, 4(1):65, 2021

    Ravi Aggarwal, Viknesh Sounderajah, Guy Martin, Daniel SW Ting, Alan Karthike- salingam, Dominic King, Hutan Ashrafian, and Ara Darzi. Diagnostic accuracy of deep learning in medical imaging: a systematic review and meta-analysis.NPJ digital medicine, 4(1):65, 2021

    2021

  2. [2]

    Dual-path frequency discriminators for few-shot anomaly detection

    Yuhu Bai, Jiangning Zhang, Zhaofeng Chen, Yuhang Dong, Yunkang Cao, and Guanzhong Tian. Dual-path frequency discriminators for few-shot anomaly detection. Knowledge-Based Systems, 302:112397, 2024

    2024

  3. [3]

    The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification, 2021

    Ujjwal Baid, Satyam Ghodasara, Michel Bilello, and et al. The rsna-asnr-miccai brats 2021 benchmark on brain tumor segmentation and radiogenomic classification, 2021

    2021

  4. [4]

    Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features.Scientific Data, 4(1):170117, 2017

    Spyridon Bakas, Hamed Akbari, and et al. Advancing the cancer genome atlas glioma mri collections with expert segmentation labels and radiomic features.Scientific Data, 4(1):170117, 2017. ISSN 2052-4463

    2017

  5. [5]

    Bmad: Benchmarks for medical anomaly detection

    Jinan Bao, Hanshi Sun, Hanqiu Deng, Yinsheng He, Zhaoxiang Zhang, and Xingyu Li. Bmad: Benchmarks for medical anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4042–4053, 2024

    2024

  6. [6]

    Improving unsupervised defect segmentation by applying structural similarity to au- toencoders.arXiv preprint arXiv:1807.02011, 2018

    Paul Bergmann, Sindy Löwe, Michael Fauser, David Sattlegger, and Carsten Steger. Improving unsupervised defect segmentation by applying structural similarity to au- toencoders.arXiv preprint arXiv:1807.02011, 2018

    arXiv 2018

  7. [7]

    Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings

    Paul Bergmann, Michael Fauser, David Sattlegger, and Carsten Steger. Uninformed students: Student-teacher anomaly detection with discriminative latent embeddings. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4183–4192, 2020

    2020

  8. [8]

    The livesr tumor segmentation benchmark (lits).Medical Image Analysis, 84:102680, 2023

    Patrick Bilic and et al. The livesr tumor segmentation benchmark (lits).Medical Image Analysis, 84:102680, 2023. ISSN 1361-8415

    2023

Show all 56 references
  1. [9]

    Informative knowledge distillation for image anomaly segmentation.Knowledge-Based Systems, 248:108846, 2022

    Yunkang Cao, Qian Wan, Weiming Shen, and Liang Gao. Informative knowledge distillation for image anomaly segmentation.Knowledge-Based Systems, 248:108846, 2022

    2022

  2. [10]

    Recent advances and clinical applications of deep learning in medical image analysis.Medical image analysis, 79:102444, 2022

    Xuxin Chen, Ximin Wang, Ke Zhang, Kar-Ming Fung, Theresa C Thai, Kathleen Moore, Robert S Mannel, Hong Liu, Bin Zheng, and Yuchen Qiu. Recent advances and clinical applications of deep learning in medical image analysis.Medical image analysis, 79:102444, 2022

    2022

  3. [11]

    Unsu- pervised anomaly detection using style distillation.IEEE Access, 8:221494–221502, 2020

    Hwehee Chung, Jongho Park, Jongsoo Keum, Hongdo Ki, and Seokho Kang. Unsu- pervised anomaly detection using style distillation.IEEE Access, 8:221494–221502, 2020

    2020

  4. [12]

    Anomaly detection via reverse distillation from one- class embedding

    Hanqiu Deng and Xingyu Li. Anomaly detection via reverse distillation from one- class embedding. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9737–9746, June 2022. 16NIE ET AL.: CFR-NET

    2022

  5. [13]

    Diagnostic assessment of deep learning algo- rithms for detection of lymph node metastases in women with breast cancer.JAMA, 318(22):2199–2210, 12 2017

    Babak Ehteshami Bejnordi and et al. Diagnostic assessment of deep learning algo- rithms for detection of lymph node metastases in women with breast cancer.JAMA, 318(22):2199–2210, 12 2017

    2017

  6. [14]

    Deep learning for medical anomaly detection–a survey.ACM Computing Surveys (CSUR), 54(7):1–37, 2021

    Tharindu Fernando, Harshala Gammulle, Simon Denman, Sridha Sridharan, and Clin- ton Fookes. Deep learning for medical anomaly detection–a survey.ACM Computing Surveys (CSUR), 54(7):1–37, 2021

    2021

  7. [15]

    Knowledge distil- lation: A survey.International Journal of Computer Vision, 129(6):1789–1819, 2021

    Jianping Gou, Baosheng Yu, Stephen J Maybank, and Dacheng Tao. Knowledge distil- lation: A survey.International Journal of Computer Vision, 129(6):1789–1819, 2021

    2021

  8. [16]

    Cflow-ad: Real-time unsu- pervised anomaly detection with localization via conditional normalizing flows

    Denis Gudovskiy, Shun Ishizaka, and Kazuki Kozuka. Cflow-ad: Real-time unsu- pervised anomaly detection with localization via conditional normalizing flows. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 98–107, January 2022

    2022

  9. [17]

    Recontrast: Domain- specific anomaly detection via contrastive reconstruction.Advances in Neural Infor- mation Processing Systems, 36:10721–10740, 2023

    Jia Guo, Shuai Lu, Lize Jia, Weihang Zhang, and Huiqi Li. Recontrast: Domain- specific anomaly detection via contrastive reconstruction.Advances in Neural Infor- mation Processing Systems, 36:10721–10740, 2023

    2023

  10. [18]

    A diffusion-based framework for multi- class anomaly detection

    Haoyang He, Jiangning Zhang, Hongxu Chen, Xuhai Chen, Zhishan Li, Xu Chen, Yabiao Wang, Chengjie Wang, and Lei Xie. A diffusion-based framework for multi- class anomaly detection. InProceedings of the AAAI conference on artificial intelli- gence, volume 38, pages 8472–8480, 2024

    2024

  11. [19]

    Fusing multispectral information for retinal layer segmentation.npj Digital Medicine, 8(1):39, 2025

    Xiang He, Fuwang Wu, Kaixuan Hu, Lizhen Cui, Weiye Song, and Yi Wan. Fusing multispectral information for retinal layer segmentation.npj Digital Medicine, 8(1):39, 2025

    2025

  12. [20]

    A semantic-enhanced method based on deep svdd for pixel-wise anomaly detection

    Chuanfei Hu, Kai Chen, and Hang Shao. A semantic-enhanced method based on deep svdd for pixel-wise anomaly detection. In2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6. IEEE, 2021

    2021

  13. [21]

    Automated segmentation of macular edema in oct using deep neural networks.Medical Image Analysis, 55:216–227, 2019

    Junjie Hu, Yuanyuan Chen, and et al. Automated segmentation of macular edema in oct using deep neural networks.Medical Image Analysis, 55:216–227, 2019. ISSN 1361-8415

    2019

  14. [22]

    Adapting visual-language models for generalizable anomaly detection in med- ical images

    Chaoqin Huang, Aofan Jiang, Jinghao Feng, Ya Zhang, Xinchao Wang, and Yanfeng Wang. Adapting visual-language models for generalizable anomaly detection in med- ical images. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11375–11385, 2024

    2024

  15. [23]

    Sarfaraz Hussein, Pujan Kandel, Candice W Bolan, Michael B Wallace, and Ulas Bagci. Lung and pancreatic tumor characterization in the deep learning era: novel supervised and unsupervised learning approaches.IEEE transactions on medical imag- ing, 38(8):1777–1787, 2019

    2019

  16. [24]

    Multi-scale feature reconstruction network for industrial anomaly de- tection.Knowledge-Based Systems, 305:112650, 2024

    Ehtesham Iqbal, Samee Ullah Khan, Sajid Javed, Brain Moyo, Yahya Zweiri, and Yusra Abdulrahman. Multi-scale feature reconstruction network for industrial anomaly de- tection.Knowledge-Based Systems, 305:112650, 2024. NIE ET AL.: CFR-NET17

    2024

  17. [25]

    Aptos 2019 blindness detection.https: //kaggle.com/competitions/aptos2019-blindness-detection,

    Karthik, Maggie, and Sohier Dane. Aptos 2019 blindness detection.https: //kaggle.com/competitions/aptos2019-blindness-detection,

    2019

  18. [26]

    Kermany and et al

    Daniel S. Kermany and et al. Identifying medical diagnoses and treatable diseases by image-based deep learning.Cell, 172(5):1122–1131.e9, 2018. ISSN 0092-8674

    2018

  19. [27]

    Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge

    Bennett Landman, Zhoubing Xu, and et al. Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge. InProc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, volume 5, page 12, 2015

    2015

  20. [28]

    Multimodal industrial anomaly detection via geometric prior.IEEE Transactions on Circuits and Systems for Video Technology, pages 1–1, 2025

    Min Li, Jinghui He, Gang Li, Jiachen Li, Jin Wan, and Delong Han. Multimodal industrial anomaly detection via geometric prior.IEEE Transactions on Circuits and Systems for Video Technology, pages 1–1, 2025. doi: 10.1109/TCSVT.2025.3613708

    2025

  21. [29]

    A survey on deep learning in medical image analysis

    Geert Litjens, Thijs Kooi, Babak Ehteshami Bejnordi, Arnaud Arindra Adiyoso Setio, Francesco Ciompi, Mohsen Ghafoorian, Jeroen Awm Van Der Laak, Bram Van Gin- neken, and Clara I Sánchez. A survey on deep learning in medical image analysis. Medical image analysis, 42:60–88, 2017

    2017

  22. [30]

    Dl- sanet: A dual-path learnable structure-prior attention network for retinal layer seg- mentation.Biomedical Signal Processing and Control, 121:110250, 2026

    Enyu Liu, Muhao Xu, Haohua Yang, Bingcan Yan, Hua Wei, and Weiye Song. Dl- sanet: A dual-path learnable structure-prior attention network for retinal layer seg- mentation.Biomedical Signal Processing and Control, 121:110250, 2026. ISSN 1746-8094. doi: https://doi.org/10.1016/j...

    2026

  23. [31]

    Unlocking the potential of reverse distillation for anomaly detection

    Xinyue Liu, Jianyuan Wang, Biao Leng, and Shuo Zhang. Unlocking the potential of reverse distillation for anomaly detection. InProceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 5640–5648, 2025

    2025

  24. [32]

    Simplenet: A simple net- work for image anomaly detection and localization

    Zhikang Liu, Yiming Zhou, Yuansheng Xu, and Zilei Wang. Simplenet: A simple net- work for image anomaly detection and localization. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20402–20411, 2023

    2023

  25. [33]

    Wei Luo, Peng Xing, Yunkang Cao, Haiming Yao, Weiming Shen, and Zechao Li. Ura- net: Uncertainty-integrated anomaly perception and restoration attention network for unsupervised anomaly detection.IEEE Transactions on Circuits and Systems for Video Technology, PP

  26. [34]

    Integrating local and global correlations with mamba-transformer for multi-class anomaly detection.Knowledge- Based Systems, 324:113740, 2025

    Zeqi Ma, Jiaxing Li, Kaihang Jiang, and Wai Keung Wong. Integrating local and global correlations with mamba-transformer for multi-class anomaly detection.Knowledge- Based Systems, 324:113740, 2025. ISSN 0950-7051. doi: https://doi.org/10.1016/ j.knosys.2025.113740. URLhttps:/...

    2025

  27. [35]

    Mocca: Multilayer one-class classification for anomaly detection.IEEE transactions on neural networks and learning systems, 33 (6):2313–2323, 2021

    Fabio Valerio Massoli, Fabrizio Falchi, Alperen Kantarci, ¸ Seymanur Akti, Hazim Ke- mal Ekenel, and Giuseppe Amato. Mocca: Multilayer one-class classification for anomaly detection.IEEE transactions on neural networks and learning systems, 33 (6):2313–2323, 2021. 18NIE ET AL....

    2021

  28. [36]

    Menze, Andras Jakab, and et al

    Bjoern H. Menze, Andras Jakab, and et al. The multimodal brain tumor image segmen- tation benchmark (brats).IEEE Transactions on Medical Imaging, 34(10):1993–2024, 2015

    1993

  29. [37]

    Few-shot medical anomaly detection through centroid consultation back and test-time self-calibration.Pattern Recognition, page 113261, 2026

    Zihan Nie, Muhao Xu, Yuan Cui, Hua Wei, Wei Yi, Sijie Niu, Yi Wan, Xunbin Wei, and Weiye Song. Few-shot medical anomaly detection through centroid consultation back and test-time self-calibration.Pattern Recognition, page 113261, 2026

    2026

  30. [38]

    Towards total recall in industrial anomaly detection

    Karsten Roth, Latha Pemula, Joaquin Zepeda, Bernhard Schölkopf, Thomas Brox, and Peter Gehler. Towards total recall in industrial anomaly detection. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14318– 14328, 2022

    2022

  31. [39]

    Asymmet- ric student-teacher networks for industrial anomaly detection

    Marco Rudolph, Tom Wehrbein, Bodo Rosenhahn, and Bastian Wandt. Asymmet- ric student-teacher networks for industrial anomaly detection. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2592–2602, 2023

    2023

  32. [40]

    Multireso- lution knowledge distillation for anomaly detection

    Mohammadreza Salehi, Niousha Sadjadi, Soroosh Baselizadeh, and et al. Multireso- lution knowledge distillation for anomaly detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14902–14912, June 2021

    2021

  33. [41]

    Deep learning in medical image analysis.Annual review of biomedical engineering, 19(1):221–248, 2017

    Dinggang Shen, Guorong Wu, and Heung-Il Suk. Deep learning in medical image analysis.Annual review of biomedical engineering, 19(1):221–248, 2017

    2017

  34. [42]

    Learn- ing and evaluating representations for deep one-class classification.arXiv preprint arXiv:2011.02578, 2020

    Kihyuk Sohn, Chun-Liang Li, Jinsung Yoon, Minho Jin, and Tomas Pfister. Learn- ing and evaluating representations for deep one-class classification.arXiv preprint arXiv:2011.02578, 2020

    2011 arXiv

  35. [43]

    Anoseg: Anomaly segmentation network using self-supervised learning.arXiv preprint arXiv:2110.03396, 2021

    Jouwon Song, Kyeongbo Kong, Ye-In Park, Seong-Gyun Kim, and Suk-Ju Kang. Anoseg: Anomaly segmentation network using self-supervised learning.arXiv preprint arXiv:2110.03396, 2021

    2021 arXiv

  36. [44]

    Does knowledge distillation really work?Advances in neural in- formation processing systems, 34:6906–6919, 2021

    Samuel Stanton, Pavel Izmailov, Polina Kirichenko, Alexander A Alemi, and An- drew G Wilson. Does knowledge distillation really work?Advances in neural in- formation processing systems, 34:6906–6919, 2021

    2021

  37. [45]

    Unsupervised visual defect detection with score-based generative model.arXiv preprint arXiv:2211.16092, 2022

    Yapeng Teng, Haoyang Li, Fuzhen Cai, Ming Shao, and Siyu Xia. Unsupervised visual defect detection with score-based generative model.arXiv preprint arXiv:2211.16092, 2022

    2022 arXiv

  38. [46]

    Two-stage reverse knowledge distilla- tion incorporated and self-supervised masking strategy for industrial anomaly detec- tion.Knowledge-Based Systems, 273:110611, 2023

    Guoxiang Tong, Quanquan Li, and Yan Song. Two-stage reverse knowledge distilla- tion incorporated and self-supervised masking strategy for industrial anomaly detec- tion.Knowledge-Based Systems, 273:110611, 2023. ISSN 0950-7051. doi: https: //doi.org/10.1016/j.knosys.2023.1106...

    2023

  39. [47]

    Anomaly detection in medi- cal imaging-a mini review

    Maximilian E Tschuchnig and Michael Gadermayr. Anomaly detection in medi- cal imaging-a mini review. InInternational Data Science Conference, pages 33–38. Springer, 2021. NIE ET AL.: CFR-NET19

    2021

  40. [48]

    Student-teacher feature pyramid matching for anomaly detection.arXiv preprint arXiv:2103.04257, 2021

    Guodong Wang, Shumin Han, Errui Ding, and Di Huang. Student-teacher feature pyramid matching for anomaly detection.arXiv preprint arXiv:2103.04257, 2021

    2021 arXiv

  41. [49]

    Tinyvit: Fast pretraining distillation for small vision transformers

    Kan Wu, Jinnian Zhang, Houwen Peng, Mengchen Liu, Bin Xiao, Jianlong Fu, and Lu Yuan. Tinyvit: Fast pretraining distillation for small vision transformers. InEuro- pean conference on computer vision, pages 68–85. Springer, 2022

    2022

  42. [50]

    Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise

    Julian Wyatt, Adam Leach, Sebastian M Schmon, and Chris G Willcocks. Anoddpm: Anomaly detection with denoising diffusion probabilistic models using simplex noise. InProceedings of the IEEE/CVF conference on computer vision and pattern recogni- tion, pages 650–656, 2022

    2022

  43. [51]

    Normal image guided segmen- tation framework for unsupervised anomaly detection.IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4639–4652, 2024

    Peng Xing, Yanpeng Sun, Dan Zeng, and Zechao Li. Normal image guided segmen- tation framework for unsupervised anomaly detection.IEEE Transactions on Circuits and Systems for Video Technology, 34(6):4639–4652, 2024. doi: 10.1109/TCSVT. 2023.3327448

    2024

  44. [52]

    Beyond feature mapping: Dual-heterogeneous knowledge distillation with mamba for industrial anomaly detection.Expert Systems with Applications, page 131146, 2026

    Muhao Xu, Zihan Nie, Baochen Fu, Zhuangzhuang Chen, Zijian Li, Hua Wei, Yi Wan, and Weiye Song. Beyond feature mapping: Dual-heterogeneous knowledge distillation with mamba for industrial anomaly detection.Expert Systems with Applications, page 131146, 2026

    2026

  45. [53]

    Reconstructed student-teacher and discriminative networks for anomaly detection

    Shinji Yamada, Satoshi Kamiya, and Kazuhiro Hotta. Reconstructed student-teacher and discriminative networks for anomaly detection. In2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 2725–2732. IEEE, 2022

    2022

  46. [54]

    Learning semantic context from normal samples for unsupervised anomaly detection

    Xudong Yan, Huaidong Zhang, Xuemiao Xu, Xiaowei Hu, and Pheng-Ann Heng. Learning semantic context from normal samples for unsupervised anomaly detection. InProceedings of the AAAI conference on artificial intelligence, volume 35, pages 3110–3118, 2021

    2021

  47. [55]

    Self-supervised learning for anomaly detection with dynamic local augmentation.IEEE Access, 9: 147201–147211, 2021

    Seungdong Yoa, Seungjun Lee, Chiyoon Kim, and Hyunwoo J Kim. Self-supervised learning for anomaly detection with dynamic local augmentation.IEEE Access, 9: 147201–147211, 2021

    2021

  48. [56]

    Msflow: Multiscale flow-based framework for unsupervised anomaly detection.IEEE Transac- tions on Neural Networks and Learning Systems, 2024

    Yixuan Zhou, Xing Xu, Jingkuan Song, Fumin Shen, and Heng Tao Shen. Msflow: Multiscale flow-based framework for unsupervised anomaly detection.IEEE Transac- tions on Neural Networks and Learning Systems, 2024

    2024

Pith tools

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