Source-linked AI summary
Subtraction-Based Tumor Segmentation and Lesion-Centered pCR Prediction for the MAMA-MIA Challenge
Kai Geissler, Raphael Schäfer
TL;DR
The paper addresses robust breast MRI tumor segmentation and pCR prediction under external multi-country evaluation. It uses subtraction-based nnU-Net segmentation and lesion-centered ensembles of pretrained 3D video classifiers, ranking second in both tasks. Segmentation generalized strongly, whereas baseline DCE-MRI pCR prediction remained limited.
Problem
The MAMA-MIA Challenge evaluates whether breast MRI models can generalize fairly across unseen sites for tumor segmentation and pCR prediction.
Method
FME uses a 5-fold nnU-Net ensemble on the first subtraction image for segmentation and 25 pretrained 3D video classifiers on lesion-centered first-three-volume crops for pCR prediction.
Results
FME ranked second in both tasks, achieving a segmentation combined score of 0.8820 and a classification combined score of 0.6642.
Takeaways & Limitations
Segmentation generalized well with robust preprocessing and strong defaults, while baseline-examination-only pCR prediction remained data- and modality-limited.
Takeaways & Limitations
pCR prediction remained close to chance, with balanced accuracy of 0.5405 only marginally different from random prediction and missing breast-density metadata limiting fairness optimization.
Abstract
from arXiv · showhide
We describe the submission of team FME to the MAMA-MIA Challenge, which evaluated primary tumor segmentation and prediction of pathological complete response (pCR) from pretreatment dynamic contrast-enhanced breast MRI on an external multi-country cohort. For segmentation, we trained a five-fold residual-encoder nnU-Net ensemble using only the first post-contrast minus pre-contrast image, combined with mirroring test-time augmentation and largest-connected-component filtering. For pCR prediction, we ensembled 25 pretrained 3D video classifiers trained on lesion-centred crops from the pre-contrast and first two post-contrast volumes. FME ranked second in both tasks. The segmentation method achieved a combined performance-fairness score of 0.882, with Dice 0.713 and normalized Hausdorff distance 0.099. The pCR method achieved a combined score of 0.664, balanced accuracy of 0.541, and equalized-odds disparity of 0.212. The results indicate that subtraction-based input and ensembling support robust tumor segmentation under cross-site domain shift, whereas pCR prediction from baseline DCE-MRI alone remains limited. For the submission repository, see https://github.com/FraunhoferMEVIS/MAMA-MIA-Challenge-FME
1 Introduction
The MAMA-MIA Challenge evaluates breast MRI tumor segmentation and pCR prediction on a hidden, multi-country cohort, emphasizing generalizability and fairness. Team FME describes its modeling choices and official results for both tasks.
- The challenge targets primary tumor segmentation and prediction of pathological complete response to neoadjuvant chemotherapy.
- Models were developed on public training data and evaluated on private test data from previously unseen sites.
- FME placed second in both tasks using subtraction-based segmentation and lesion-centered pCR classification pipelines.
- The segmentation pipeline produced accurate, stable masks under domain shifts with competitive fairness, while pCR prediction achieved balanced accuracy above baseline despite low task-wide performance.
- The study details the datasets, challenge metrics, modeling choices, and official results.
2 The MAMA-MIA challenge
MAMA-MIA combines public multi-center training data with a hidden multi-country test cohort and ranks methods using equally weighted predictive performance and subgroup fairness. Fairness is assessed across age, menopausal status, and breast density using task-specific metrics.
- Datasets: The challenge uses a public training dataset and a private hidden test dataset from Poland, Lithuania, and Spain.The private test set contains 574 cases across three institutions.
- Datasets: The public training cohort contains 1,506 MRI volumes from 25 US clinical centers, with expert primary-tumor annotations and pCR labels for most cases.
- Challenge metrics: The final ranking score equally weights predictive performance and subgroup fairness.
- Challenge metrics: Fairness is evaluated across age, menopausal status, and breast density using differences between maximum and minimum subgroup-average metric values.
- Challenge metrics: Segmentation performance combines Dice similarity coefficient and normalized Hausdorff distance, while classification performance uses balanced accuracy.
- Challenge metrics: Classification fairness is reported through equalized-odds disparity, for which lower leaderboard values are better.
- Submitted pipelines: The segmentation approach uses a 5-fold nnU-Net ensemble on the first subtraction volume, whereas pCR prediction uses a 25-fold video-classifier ensemble on lesion-cropped first-three-timepoint inputs.
3 Methods
The methods use subtraction-based, ensembled nnU-Net segmentation and lesion-centered 3D video classification from the first three DCE-MRI volumes. Cross-validation, augmentation, normalization, cropping, and ensemble averaging support inference for both tasks.
- Segmentation: The segmentation input is the first post-contrast minus pre-contrast image, chosen to reduce variation from site-specific acquisition timing.
- Segmentation: Segmentation uses a 5-fold nnU-Net ensemble with framework-based resampling and intensity normalization.
- Segmentation: Mirroring test-time augmentation and largest-connected-component filtering are applied during segmentation inference.
- Classification: The pCR classifier receives the first three DCE-MRI volumes so pretrained video models can use three input channels.
- Classification: Lesion bounding boxes are cropped using reference segmentations during training and predicted masks during inference, then resampled to model-specific fixed sizes.
- Classification: Hyperparameter search with 5-fold cross-validation selected five configurations across five folds, producing a final ensemble of 25 pretrained TorchVision video models.
- Classification: Classification inference averages ensemble and axis-flipping test-time-augmentation predictions using a 0.5 threshold.
4 Results
Team FME ranked second in both MAMA-MIA tasks, with closely clustered segmentation results and classification performance above the baseline but limited fairness.
- Segmentation: 0.8820 combined score placed FME second in segmentation, behind 0.8858 and ahead of 0.8782.FME’s fairness score was 0.9574 and performance score was 0.8066.
- Segmentation: 0.7125 Dice and 0.0993 normalized Hausdorff distance characterized FME’s second-place segmentation result.The top three normalized Hausdorff distances were tightly clustered at 0.0990, 0.0993, and 0.1017.
- Classification: 0.6642 combined score placed FME second in classification among 15 submitted teams.Only three teams outperformed the random classifier baseline with respect to combined score.
- Classification: 0.5405 balanced accuracy and 0.2121 equalized-odds disparity were reported for FME’s classification submission.The corresponding top-place values were 0.4964 balanced accuracy and 0.1150 equalized-odds disparity.
5 Discussion
The discussion attributes robust segmentation to subtraction-based inputs, ensembling, and conservative post-processing, while identifying persistent limits for fairness and pCR prediction.
- Segmentation: A five-fold nnU-Net ensemble using only the first subtraction image and connected-component post-processing remained robust across hidden multi-country test sites.The authors hypothesize that the subtraction input mitigated acquisition-timing variability.
- Segmentation: Largest-component filtering reduced small spurious detections and improved boundary metrics, but may undersegment multifocal disease.The authors report that this trade-off did not harm segmentation performance in the challenge setting.
- Segmentation: Using only the first post-contrast minus pre-contrast image simplified modeling and avoided site-specific multi-phase timing differences.This design emphasized lesion morphology, spatial localization, and boundary delineation rather than kinetic curve modeling.
- Fairness: 0.934–0.962 fairness scores were close among the top 12 segmentation methods, while performance scores varied from 0.736–0.819.FME did not explicitly optimize fairness beyond robustness-oriented measures.
- Classification: 0.5405 balanced accuracy was only marginally different from random prediction for pCR classification, with p = 0.063.The authors report that baseline DCE-MRI provides weak signal without richer imaging or clinical covariates.
- Classification: Similar performance across 2D, 2D-with-3D-attention, and 3D classifiers suggested that hyperparameter and ensemble improvements did not overcome fundamental signal constraints.Missing breast-density metadata also limited fairness optimization.
6 Conclusion
Team FME placed second in both challenge tasks, revealing a practical contrast between robust segmentation and data- and modality-limited pCR prediction.
- Conclusion: Team FME placed second in both segmentation and classification and was the only submission in the top three for both tasks.The conclusion contrasts robust preprocessing and strong defaults for segmentation with limitations in baseline-examination-only pCR prediction.
- Conclusion: Future progress may depend on comprehensive imaging, additional clinical inputs, calibrated decision policies, and fairness-aware training supported by complete metadata.The conclusion identifies these directions rather than favoring classifier architecture alone.