Mask-Guided Multi-Channel SwinUNETR Framework for Robust MRI Classification

Karim Lekadir; Lidia Garrucho; Oliver Diaz; Richard Osuala; Smriti Joshi

Mask-Guided Multi-Channel SwinUNETR Framework for Robust MRI Classification

T0 review · 0 major / 5 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 The paper reports that a breast-mask-guided, multi-channel SwinUNETR ensemble achieves a held-out AUC of 0.86 in the ODELIA breast MRI challenge, ranking second.

desk verdict Solid externally validated challenge result—the leaderboard AUC is credible—but the causal claims about masking and ensembling outrun the evidence. read the letter →

arxiv 2508.20621 v1 pith:BNKS2TUI submitted 2025-08-28 cs.CV

Smriti Joshi , Lidia Garrucho , Richard Osuala , Oliver Diaz , Karim Lekadir This is my paper · ORCID

classification cs.CV

keywords breastcancerMRIODELIAchallengeSwinUNETRdynamiccontrast-enhancedmaximumintensityprojectionensemblelearningclassimbalancemulti-centergeneralizability

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's aim is to show that a tissue-focused, well-engineered deep pipeline generalizes across the scanner and center heterogeneity of the ODELIA breast MRI challenge. Its recipe is to mask the breast region before computing input images, stack four DCE-MRI projections that encode early enhancement and washout, train a SwinUNETR classifier with heavy augmentation, and ensemble models trained with and without inverse-frequency class weighting. On the challenge's held-out test set, the ensemble reached AUC 0.86, sensitivity 0.62 at 90% specificity, and specificity 0.57 at 90% sensitivity, and the method placed second on the leaderboard. The authors argue this supports automated breast MRI triage; they also note that benign versus normal overlaps remain the main failure mode.

What carries the argument

The load-bearing object is the four-channel breast-masked MIP input: a fixed set of maximum intensity projections computed inside an automatically segmented breast mask, comprising the first post-contrast phase, the first two subtractions from the pre-contrast baseline, and the last available subtraction. The classification backbone is SwinUNETR, a transformer-based encoder-decoder that pairs shifted-window self-attention with U-Net-style down/up sampling; a lightweight head pools its encoder features into per-breast class probabilities. Two models with different loss weighting—natural class weights and inverse-frequency weights—are trained, and their probabilities are averaged into the fina

What would settle it

Retrain the classifier on ODELIA data with (a) no masking, (b) predicted masks, and (c) manual ground-truth masks for a subset, then compare held-out AUC; also compute Dice scores between predicted and manual masks on ODELIA studies. Masking is a causal contributor only if predicted-mask AUC exceeds no-mask AUC and tracks mask quality.

Watch

Extended reading notes

Core claim

The central claim is that representing each breast as four masked MIP channels—first post-contrast, subtraction 1, subtraction 2, and the last subtraction—preserves the contrast-uptake kinetics that separate benign and malignant lesions, and that the SwinUNETR backbone can exploit those kinetics once background signal is removed. The paper further claims that two class-imbalance strategies are complementary: natural weighting preserves overall AUC by favoring the majority normal class, while inverse-frequency weighting improves benign detection; ensembling their probabilities recovers sensitivity without collapsing specificity. Ensembling across five patient-stratified folds, the model reach

Load-bearing premise

The load-bearing premise is that the breast-mask segmenter, trained on 36 Duke and 100 Yunnan cases, yields accurate masks on ODELIA images from six European centers with different scanners and protocols; the paper reports no mask-quality metric on ODELIA data and no ablation that removes masking.

Editorial extensions

If this is right

The combination of breast masking, four MIP channels, and weighted-loss ensembling is a transferable recipe for multi-center DCE-MRI classification, and the authors release the code publicly.
Including the first post-contrast phase alongside subtraction images avoids disadvantaging lesions already visible in pre-contrast or fat-suppressed images.
Ensembling natural and inverse-frequency weighted losses balances overall AUC with minority-class sensitivity; models without weighting skew toward the majority normal class.
The held-out test results constitute evidence that a single model can handle multi-vendor, 1.5 T and 3 T data without per-center adaptation.
Benign-versus-normal misclassifications dominate, indicating that contrast kinetics alone may be insufficient for this boundary.

Reading between the lines

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

The paper's assumption that an externally trained breast segmenter transfers to six European centers is unmeasured; a Dice-score audit on ODELIA masks or an ablation that removes masking would isolate how much of the AUC 0.86 comes from masking versus the backbone and ensemble.
The left/right breast split at 50% image width presumes symmetric anatomy and consistent patient centering; on off-center acquisitions, lesions near the midline or far lateral could be assigned to the wrong breast half.
Because the 'last post-contrast' channel is per-case (between two and seven phases), the fourth channel mixes different kinetic time points; fixing a canonical late phase or adding the phase count as a covariate might further stabilize cross-center performance.
The framework could be extended to include T2-weighted or diffusion-weighted sequences as additional channels, which the authors mention as future work but do not test.

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.

Desk Editor's Note

read the letter

Short version: this is a solid challenge report with one externally grounded result—second place on ODELIA, test AUC 0.8610—and a few design claims that don't yet have evidence behind them. Worth engaging.

What's actually new: the specific test-set performance of this pipeline on the ODELIA held-out set. The components (SwinUNETR, MIP inputs, class-weighted loss, ensembling, augmentation) are all known and properly cited; the contribution is a reproducible multi-center benchmark baseline, not a new mechanism.

What it does well: the central evaluation has external teeth. The test AUC and rank come from the challenge organizers, not the authors. CV is patient-stratified, code is public, and the preprocessing details (resampling, normalization stats, augmentation, training schedule) are specific enough to reproduce. The failure-case analysis is honest about normal/benign overlap, and the class-weighting tradeoff is clearly described.

Soft spots, in proportion. First, no ablations for the two choices the Discussion credits: masking and ensembling. The claim that segmentation-guided classification 'reduced noise and improved discriminative power' is unsupported without a with/without-mask comparison, and the ensemble's contribution to the test score is unknown because test-set metrics for the individual fold models aren't reported. Second, no mask-quality metric on ODELIA data. If the Duke/Yunnan segmentation model mis-segments the six European centers, the masked inputs could be excluding or leaking tissue. That doesn't invalidate the reported AUC—the submitted predictions were what they were—but it does undermine the causal story. Third, Table 1 is confusingly labeled: the per-fold rows are validation results, not test, and only the final row is the held-out test. Minor, but should be fixed.

The stress-test is correct: no significant objection to the central empirical claim. The missing ablations weaken the explanation, not the leaderboard number.

Who this is for: researchers looking for an open baseline in multi-center breast MRI classification, and anyone following ODELIA. It deserves a serious referee. Send it to review, with a request for ablations (or softened causal claims) and a table fix. I wouldn't cite it in my own work in the next year unless I moved into breast MRI, but it's a reasonable citation for those who do.

Referee Report

0 major / 5 minor

Summary. The paper presents a deep learning framework for the ODELIA challenge task of classifying each breast in DCE-MRI as normal, benign, or malignant. The method uses four-channel maximum intensity projections (first post-contrast, first/second subtraction, last subtraction) from masked breast volumes, a SwinUNETR backbone with a lightweight classification head, inverse-frequency class weighting, and ensembling across five patient-stratified cross-validation folds. The authors report a held-out test-set AUC of 0.8610 and an overall score of 0.6830, which they state placed second on the challenge leaderboard. The code is publicly available.

Significance. The central empirical claim is externally grounded: the test-set metrics were computed by the ODELIA challenge organizers on a held-out set, and the reported ensemble score is arithmetically consistent with the challenge's stated metric (average of AUC, sensitivity at 90% specificity, and specificity at 90% sensitivity). The cross-validation is patient-stratified, which avoids the most common leakage failure. Public code is a strength. The main weakness is that causal claims about the mask-guided and multi-channel design are not backed by ablations, and no mask-quality metric on ODELIA data is reported. These gaps affect the paper's explanatory narrative about why the method works, but they do not undermine the externally validated test-set result.

minor comments (5)

[§4 Discussion] The sentence 'The integration of a segmentation-guided classification strategy provided a tissue-focused representation that excluded background signal, reducing noise and improving discriminative power' is a causal claim that is not supported by any ablation or mask-quality evaluation. The segmentation model was trained on only 136 cases from Duke and Yunnan and applied to six European centers; no metric or visual check on ODELIA masks is provided. Please either add a no-mask baseline on the validation folds and a mask-overlap statistic, or temper the wording to 'may have contributed' consistently with the earlier hedging.
[§3.4, Table 1] The held-out test result is reported only for the ensemble across all folds. Please specify how the ensemble was formed (e.g., averaging probabilities of five fold models) and whether per-fold test predictions were available but not reported. If only one official submission was allowed, state this explicitly. Reporting per-fold test scores, or at least a confidence interval via bootstrapping on the test predictions, would strengthen the result.
[§3.3 Implementation Details] The augmentation and normalization steps are described without specifying whether they are applied to the 3D volumes before computing MIPs or to the 2D MIP inputs. This matters for reproducibility, especially because breast masks are 3D and the MIP projection is computed along z. Please clarify the order of operations and whether the mask is resampled to the same 0.7×0.7×3 mm grid.
[Table 1] The first column header 'Test Set Weights' is confusing: the rows are fold indices and weight schemes, while only the last row is actually the held-out test set. Rename the column (e.g., 'Configuration') and label rows 'Fold 1, no weighting' / 'Fold 1, weighted' / 'Held-out ensemble'. Also, the text in §3.4 refers to 'micro-AUC' while the table header says 'AUC'; please reconcile the terminology.
[Throughout] Minor editorial issues: 'to ensurespatial' should be 'to ensure spatial'; 'dependencies MRI data' appears to be missing 'in' ('dependencies in MRI data'); 'T able 1' has an extra space; in Figure 4 there is an extra space before the comma after 'Figure 4e.)'. These do not affect the content.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the leaderboard claim rests on externally scored held-out test predictions; preprocessing and ensembling choices are standard and do not encode test labels.

full rationale

The central claim is the reported 0.8610 test AUC and second-place rank on the ODELIA held-out challenge test set (Section 3.4, Table 1). This is an external benchmark evaluation: the challenge organizers computed the metrics on a held-out test set from the submitted predictions, and no parameter in the paper is fitted to those test labels. The breast-mask segmentation model is trained on separate public datasets (Duke, Yunnan; Section 3.1), so the masking preprocessing does not incorporate ODELIA test information. The choice of class weighting and ensembling was made using five-fold cross-validation on the training portion, which is standard practice and does not make the test result circular. The only self-citations (references [3] and [6]) are related-work pointers in the Discussion and are not load-bearing for any derivation. The Discussion's causal claim that masking 'improved discriminative power' lacks an ablation, but that is a support gap, not circularity. No step in the derivation reduces by construction to its own inputs, and no fitted parameter is renamed as a prediction.

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

The framework is an empirical pipeline built from standard deep-learning components; it introduces no new mathematical entities or postulates. Its assumptions are domain-level heuristics about breast MRI kinetics and transferability of the mask model.

free parameters (3)

Channel-wise normalization statistics = means (0.2074, 0.1290, 0.1396, 0.1470), stds (0.2110, 0.1629, 0.1620, 0.1626)
Listed in Section 3.3; computed from the training data and would shift on different populations.
Resampling grid and crop size = 512x512x32 voxels, height cropped to 256
Chosen to standardize heterogeneous acquisitions; the fixed 32-slice depth may discard anatomy in cases with more slices.
Training hyperparameters = lr 1e-4, segmentation 100 epochs, classification 300 epochs, batch 20/10
Chosen by hand; no sensitivity analysis reported.

assumptions (3)

domain assumption Malignant lesions typically show rapid initial enhancement and washout whereas benign lesions show slower progressive enhancement
Section 2.1 cites this clinical heuristic to justify the choice of MIP channels; if DCE kinetic patterns do not separate classes in the ODELIA population, the input design loses its rationale.
domain assumption The externally trained breast segmentation masks are accurate enough on ODELIA images to exclude background without removing lesion tissue
Section 2.1 and 3.1; the segmentation model is trained on 136 cases from Duke and Yunnan, and no segmentation accuracy on ODELIA data is reported.
domain assumption MIP projection along z preserves diagnostically relevant enhancement kinetics
Section 2.1; projecting volumes to 2D MIPs collapses spatial information, and the paper does not compare MIPs to volume-based or slice-based inputs.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Mask-Guided Multi-Channel SwinUNETR Framework for Robust MRI Classification." pith.science (2026). https://pith.science/paper/BNKS2TUI

@misc{pith2026250820621,
  author       = {Pith},
  title        = {Pith review of: Mask-Guided Multi-Channel SwinUNETR Framework for Robust MRI Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BNKS2TUI}},
  note         = {Machine review of arXiv:2508.20621}
}

read the original abstract

Breast cancer is one of the leading causes of cancer-related mortality in women, and early detection is essential for improving outcomes. Magnetic resonance imaging (MRI) is a highly sensitive tool for breast cancer detection, particularly in women at high risk or with dense breast tissue, where mammography is less effective. The ODELIA consortium organized a multi-center challenge to foster AI-based solutions for breast cancer diagnosis and classification. The dataset included 511 studies from six European centers, acquired on scanners from multiple vendors at both 1.5 T and 3 T. Each study was labeled for the left and right breast as no lesion, benign lesion, or malignant lesion. We developed a SwinUNETR-based deep learning framework that incorporates breast region masking, extensive data augmentation, and ensemble learning to improve robustness and generalizability. Our method achieved second place on the challenge leaderboard, highlighting its potential to support clinical breast MRI interpretation. We publicly share our codebase at https://github.com/smriti-joshi/bcnaim-odelia-challenge.git.

Figures

Figures reproduced from arXiv: 2508.20621 by the authors.

**Figure 1.** Representative examples of maximum intensity projections (MIPs) showing contrast enhancement kinetics in breast MRI. Phase 1 corresponds to the first postcontrast phase of the DCE-MRI. Subtraction images (subx) denote the x th postcontrast phase minus the pre-contrast phase. The last phase is dynamically selected for each case, as the number of post-contrast phases varies. The left panel demonstrates a malignant l… view at source ↗

**Figure 2.** Overview of the ODELIA challenge dataset. Left: Contribution of different centers to the total dataset, illustrating the multi-center and heterogeneous nature of the data. Right: Class distribution across all breast labels, showing the relative proportions of “no lesion,” “benign lesion,” and “malignant lesion.” Abbreviations: CAM: Cambridge University Hospitals, Cambridge, UK; MHA: Mitera Hospital, Athens, Greece; … view at source ↗

**Figure 3.** Confusion matrices comparing classification performance with natural class weighting, inverse-frequency class weighting, and their ensemble. Class weighting improves benign detection but increases false positives for normal cases. Ensembles strikes a balance between the two approaches. without weighting achieved higher micro-AUC by favoring the majority class, at the cost of reduced sensitivity for benign and malig… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Examples of misclassified cases. The ground-truth label is overlaid on each corresponding image. benign and malignant tumors. The integration of a segmentation-guided classification strategy provided a tissue-focused representation that excluded background signal, re…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

14 extracted references · 9 canonical work pages

[1]

Information11(2), 125 (2020)

Buslaev, A., Iglovikov, V.I., Khvedchenya, E., Parinov, A., Druzhinin, M., Kalinin, A.A.: Albumentations: fast and flexible image augmentations. Information11(2), 125 (2020)

work page 2020
[2]

arXiv preprint arXiv:2211.02701 (2022)

Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al.: Monai: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 (2022)

arXiv 2022
[3]

Journal of Imaging 11(4), 98 (2025)

Dimitriadis, A., Kalliatakis, G., Osuala, R., Kessler, D., Mazzetti, S., Regge, D., Diaz, O., Lekadir, K., Fotiadis, D., Tsiknakis, M., et al.: Assessing cancer presence in prostate mri using multi-encoder cross-attention networks. Journal of Imaging 11(4), 98 (2025)

work page 2025
[4]

https://gco.iarc.fr/ (2023), https://gco.iarc.fr/, accessed: 2023- 08-07

Global Cancer Observatory: The global cancer observatory (gco) is an interactive web-based platform presenting global cancer statistics to inform cancer control and research. https://gco.iarc.fr/ (2023), https://gco.iarc.fr/, accessed: 2023- 08-07

work page 2023
[5]

In: International MICCAI brainlesion workshop

Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In: International MICCAI brainlesion workshop. pp. 272–284. Springer (2021)

2021
[6]

Joshi, S., Osuala, R., Garrucho, L., Kushibar, K., Kessler, D., Diaz, O., Lekadir, K.: Single image test-time adaptation via multi-view co-training (2025),https: //arxiv.org/abs/2506.23705

work page Pith review arXiv 2025
[7]

Nature reviews Clinical oncology 14(12), 749–762 (2017)

Lambin, P., Leijenaar, R.T., Deist, T.M., Peerlings, J., De Jong, E.E., Van Tim- meren, J., Sanduleanu, S., Larue, R.T., Even, A.J., Jochems, A., et al.: Radiomics: the bridge between medical imaging and personalized medicine. Nature reviews Clinical oncology 14(12), 749–762 (2017)

work page 2017
[8]

Radiographics 26(6), 1719–1734 (2006)

Macura, K.J., Ouwerkerk, R., Jacobs, M.A., Bluemke, D.A.: Patterns of enhance- ment on breast mr images: interpretation and imaging pitfalls. Radiographics 26(6), 1719–1734 (2006)

work page 2006

Show all 14 references

[9]

Müller-Franzes, G., Sánchez, L.E., Payne, N., Athanasiou, A., Kalogeropoulos, M., Lopez, A., Busto, A.M.S., Herrero, J.C., Rasoolzadeh, N., Zhang, T., Mann, R., Jutz, D., Bode, M., Kuhl, C., Veldhuis, W., Saldanha, O.L., Zhu, J., Kather, J.N., Truhn, D., Gilbert, F.J.: A europ...

2025 arXiv
[10]

the Journal of machine Learning research12, 2825–2830 (2011)

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., et al.: Scikit-learn: Machine learning in python. the Journal of machine Learning research12, 2825–2830 (2011)

2011
[11]

Computer Methods and Programs in Biomedicine 208, 106236 (Sep 2021).https://doi.org/10.1016/j.cmpb.2021.106236, http: //dx.doi.org/10.1016/j.cmpb.2021.106236

Pérez-García, F., Sparks, R., Ourselin, S.: Torchio: A python library for effi- cient loading, preprocessing, augmentation and patch-based sampling of medi- cal images in deep learning. Computer Methods and Programs in Biomedicine 208, 106236 (Sep 2021).https://doi.org/10.1016...

2021
[12]

In: International Conference on Medical image computing and computer-assisted intervention

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomedi- cal image segmentation. In: International Conference on Medical image computing and computer-assisted intervention. pp. 234–241. Springer (2015)

2015
[13]

The Can- cer Imaging Archive (2021).https://doi.org/https://doi.org/10.7937/TCIA

Saha, A., Harowicz, M.R., Grimm, L.J., Weng, J., Cain, E.H., Kim, C.E., Ghate, S.V., Walsh, R., Mazurowski, M.A.: Dynamic contrast-enhanced magnetic reso- nance images of breast cancer patients with tumor locations [data set]. The Can- cer Imaging Archive (2021).https://doi.or...

2021 doi
[14]

Zhang, J.: Breast_Cancer DCE-MRI Data. Dataset on Zenodo (2023).https:// doi.org/10.5281/zenodo.8068383, https://doi.org/10.5281/zenodo.8068383, contains 100 DCE-MRI cases from Yunnan Cancer Hospital with both tumor and whole-breast annotations for segmentation

2023 doi

Pith tools

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

[1] [1]

Information11(2), 125 (2020)

Buslaev, A., Iglovikov, V.I., Khvedchenya, E., Parinov, A., Druzhinin, M., Kalinin, A.A.: Albumentations: fast and flexible image augmentations. Information11(2), 125 (2020)

work page 2020

[2] [2]

arXiv preprint arXiv:2211.02701 (2022)

Cardoso, M.J., Li, W., Brown, R., Ma, N., Kerfoot, E., Wang, Y., Murrey, B., Myronenko, A., Zhao, C., Yang, D., et al.: Monai: An open-source framework for deep learning in healthcare. arXiv preprint arXiv:2211.02701 (2022)

arXiv 2022

[3] [3]

Journal of Imaging 11(4), 98 (2025)

Dimitriadis, A., Kalliatakis, G., Osuala, R., Kessler, D., Mazzetti, S., Regge, D., Diaz, O., Lekadir, K., Fotiadis, D., Tsiknakis, M., et al.: Assessing cancer presence in prostate mri using multi-encoder cross-attention networks. Journal of Imaging 11(4), 98 (2025)

work page 2025

[4] [4]

https://gco.iarc.fr/ (2023), https://gco.iarc.fr/, accessed: 2023- 08-07

Global Cancer Observatory: The global cancer observatory (gco) is an interactive web-based platform presenting global cancer statistics to inform cancer control and research. https://gco.iarc.fr/ (2023), https://gco.iarc.fr/, accessed: 2023- 08-07

work page 2023

[5] [5]

In: International MICCAI brainlesion workshop

Hatamizadeh, A., Nath, V., Tang, Y., Yang, D., Roth, H.R., Xu, D.: Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images. In: International MICCAI brainlesion workshop. pp. 272–284. Springer (2021)

2021

[6] [6]

Joshi, S., Osuala, R., Garrucho, L., Kushibar, K., Kessler, D., Diaz, O., Lekadir, K.: Single image test-time adaptation via multi-view co-training (2025),https: //arxiv.org/abs/2506.23705

work page Pith review arXiv 2025

[7] [7]

Nature reviews Clinical oncology 14(12), 749–762 (2017)

Lambin, P., Leijenaar, R.T., Deist, T.M., Peerlings, J., De Jong, E.E., Van Tim- meren, J., Sanduleanu, S., Larue, R.T., Even, A.J., Jochems, A., et al.: Radiomics: the bridge between medical imaging and personalized medicine. Nature reviews Clinical oncology 14(12), 749–762 (2017)

work page 2017

[8] [8]

Radiographics 26(6), 1719–1734 (2006)

Macura, K.J., Ouwerkerk, R., Jacobs, M.A., Bluemke, D.A.: Patterns of enhance- ment on breast mr images: interpretation and imaging pitfalls. Radiographics 26(6), 1719–1734 (2006)

work page 2006