Source-linked AI summary
Feature Space Guidance for Breast Cancer Classification in DCE-MRI
Benjamin Hamm, Yannick Kirchhoff, Maximilian Rokuss, Moritz Langenberg, Constantin Ulrich, Tassilo Wald, Jeremias Traub, Karol Gotkowski, Klaus Maier-Hein
TL;DR
Automated DCE-MRI classification faces high-dimensional scans, small lesions, and cross-center protocol variation. The paper combines anatomical localization, adaptive temporal feature selection, and lesion-segmentation pretraining; it reports performance gains and first place in the MICCAI ODELIA Breast MRI Challenge 2025.
Problem
Automated classification must handle high-dimensional DCE-MRI, small lesions, and acquisition-protocol differences across centers.
Source: §1 · p. 2Method
The framework combines breast localization, adaptive selection of temporal features in latent space, and supervised lesion-segmentation pretraining for classification.
Source: §1 · p. 3, §2.1 · p. 4, §2.2 · p. 5Results
The authors report performance gains from structured, anatomy-aware, temporally adaptive design and first place in the MICCAI ODELIA Breast MRI Challenge 2025.
Source: §4 · pp. 8–9Takeaways & Limitations
The reported gains are associated with explicitly structured representation design rather than model capacity or self-supervised pretraining scale alone.
Source: §4 · p. 8
Abstract
from arXiv · showhide
Dynamic contrast enhanced breast MRI (DCE-MRI) is a powerful clinical tool for breast cancer detection, providing high resolution anatomical detail together with rich temporal contrast information. However, high dimensional 4D inputs, small lesions, and heterogeneous acquisition protocols across clinical sites hinder robust automated classification of healthy, benign, and malignant cases. To address these challenges, we propose a framework that dynamically analyzes latent representations to adapt to protocol-specific characteristics. Spatial variability is mitigated by reducing confounding background uptake and compensating for misalignment caused by deformable soft tissue. Additionally, relationships in the latent space across phases are leveraged to select the most informative temporal features, improving robustness to protocol-specific temporal variability. Finally, task specific discriminative features are promoted through large scale supervised lesion segmentation pretraining, which substantially enhances downstream finetuning. Evaluated under leave-one-center-out validation on the ODELIA dataset and the held-out AMBL cohort, the proposed framework substantially outperforms finetuned radiology foundation models and prior methods, improving mean AUROC by nearly 8 points and balanced accuracy by 4 points over the strongest baseline. Additionally, our method achieved first place in the MICCAI ODELIA Breast MRI Challenge 2025, further demonstrating its effectiveness for robust breast cancer classification. We publicly release our codebase under https://github.com/MIC-DKFZ/CURIAtor.
1 Introduction
DCE-MRI provides anatomical and temporal information for breast assessment, but automated classification must contend with high-dimensional scans, small lesions, and substantial differences between acquisition protocols. The paper introduces a framework that combines anatomical localization, adaptive temporal feature selection, and supervised pretraining to address these challenges.
1 Introduction
- DCE-MRI combines high-resolution anatomical information with temporal signatures of tissue perfusion and permeability for breast assessment.The passage notes improved sensitivity compared with conventional mammography and ultrasound, particularly in dense breast tissue.Source: §1 · p. 2
- Protocol variation, high-dimensional 4D scans, and lesions occupying a small fraction of the imaged volume make reliable automated DCE-MRI classification challenging.Differences in phase counts, temporal resolution, and contrast timing across scanners and institutions create distribution shifts.Source: §1 · p. 2
- Fixed phase selection assumes consistent temporal sampling, while maximum intensity projections discard volumetric depth and contrast-uptake dynamics.These simplifications may suppress subtle enhancement differences or allow strongly enhancing surrounding regions to dominate.Source: §1 · p. 2
- The framework addresses spatial variability through anatomical localization, temporal variability through adaptive latent-feature selection, and weak case-level supervision through anatomically grounded pretraining.The Curia radiology foundation model embeds each temporal volume, and the framework selects informative features across phases rather than using a fixed acquisition scheme.Source: §1 · p. 3
- The paper presents spatial alignment, temporal alignment, and supervised pretraining as a unified three-stage hierarchy for breast cancer classification.Source: §1 · p. 4
2 Methods
The methods build a hierarchy from breast localization to adaptive temporal representation and lesion-classification transfer learning. BreastDivider constrains analysis anatomically, CURIAtor selects a complementary phase from each acquisition, and segmentation pretraining initializes the classifier.
2 Methods
- The framework progressively abstracts anatomical, temporal, and task-specific information from heterogeneous DCE-MRI data.Source: §2 · p. 4
2.1 BreastDivider Model
- BreastDivider is a 3D segmentation model trained to localize and separate breast regions, with reported localization Dice of 99.08 ± 0.48.Its curated training resource comprises 17,956 scans from more than 55 centers and spans multiple imaging sequences.Source: §2.1 · p. 4
2.2 CURIAtor Module
- CURIAtor embeds temporal volumes, pools slice features, and uses ROI overlap weights to reduce background contributions when computing pairwise volume dissimilarities.Volumes are cropped to the breast-mask bounding box before feature extraction; the ROI mask is projected onto the token grid.Source: §2.2 · p. 5
- CURIAtor fixes Pre and Post1, then selects the most dissimilar remaining post-contrast phase to form a three-phase representation.The selection uses learned embedding distances without explicit phase labels or assumptions about regular sampling.Source: §2.2 · pp. 4–5
2.3 Transfer Learning Framework
- Breast tumor segmentation pretraining on MAMA-MIA initializes downstream classification with anatomically informed representations.The pretraining data comprise more than 17,000 three-dimensional breast volumes; the encoder is then used as a universal initialization.Source: §2.3 · pp. 5–6
3 Experiments and Results
The experiments test cross-center classification on ODELIA and held-out AMBL, with comparisons against radiology foundation models and MIP-based methods. The reported results favor the framework across centers and favor adaptive temporal selection over static configurations.
3 Experiments and Results
- The evaluation uses five-fold ODELIA cross-validation with each fold representing a distinct clinical center, plus held-out AMBL evaluation.Subgroup analyses are precluded by the lack of harmonized cohort metadata.Source: §3 · p. 6
- The framework consistently outperforms radiology foundation models across centers and improves on MIP-based approaches.The comparison covers ODELIA centers and AMBL; the supplied passages do not provide the table’s individual metric values.Source: Table 1 · p. 7, §3 · p. 7
- The benchmark reports AUROC and balanced accuracy for each ODELIA center, AMBL, and their mean, with bootstrap 95% confidence intervals.The ablation study reports mean AUROC and balanced accuracy across centers.Source: Table 1 · p. 7, Table 2 · p. 8
- Removing pretraining or replacing BreastDivider ROI prediction with whole-volume processing sharply reduces performance.The ablations assess transfer learning and anatomically constrained feature extraction.Source: §3 · p. 7, Table 2 · p. 8
- CURIAtor selection consistently outperforms static sequence configurations, supporting dynamic temporal weighting as a discriminative signal.The ablation table reports mean AUROC and balanced accuracy across centers, with relative changes measured against the final CURIAtor framework.Source: §3 · p. 8, Table 2 · p. 8
4 Discussion and Conclusion
The framework combines spatial alignment, temporal adaptation, and supervised lesion pretraining for DCE-MRI classification. Cross-center evaluation reports gains over prior baselines, and the method placed first in the MICCAI ODELIA Breast MRI Challenge 2025.
4 Discussion and Conclusion
- Mean AUROC of 75.13 and mean BA of 64.42 across centers exceed the strongest MIP baseline by nearly 8 AUROC points and 4 BA points.Replacing CURIAtor with static phase combinations reduces performance by up to 11.65 AUROC points and over 3 BA points.Source: §4 · p. 8
- The framework unifies spatial alignment, temporal alignment, and supervised lesion pretraining in a three-stage hierarchy for DCE-MRI classification.The authors describe the design as anatomy-aware and temporally adaptive.Source: §4 · p. 8
- The method achieved first place in the MICCAI ODELIA Breast MRI Challenge 2025.The authors present this as further evidence of effectiveness for robust clinical breast cancer classification.Source: §4 · p. 9