Central Asian Research Index

Scientific Journal of Astana IT University, 2025, с. 172-184

CHALLENGES IN GENERALIZING BREAST MRI TUMOR SEGMENTATION ACROSS MULTIPLE DATASETS

Beibit Abdikenov, Victor Suvorov

Аннотация · Abstract

Accurate segmentation of breast tumors in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is essential for precise diagnosis, treatment planning, and quantitative analysis. While deep learning methods have achieved strong performance in controlled research settings, their ability to generalize across diverse clinical datasets remains underexplored and poses a major barrier to clinical adoption. In this study, we evaluate the cross-dataset generalizability of a 3D Residual U-Net model using the multicenter MAMA-MIA benchmark, which consolidates four publicly available breast MRI collections annotated by expert radiologists. A leave-one-out experimental design is employed, with three datasets used for training and validation, and the remaining dataset held-out for independent testing to simulate real-world deployment scenarios. Model performance is assessed using Dice coefficient, Precision, and Recall, alongside quantitative analysis of tumor volume estimation accuracy. The best Dice score achieved by our model was 0.683 when tested on the NACT subset. Results show a consistent degradation in segmentation accuracy when models are applied to unseen datasets, indicating that performance declines significantly outside the distribution of the training data. The most pronounced drop occurs when the DUKE dataset serves as the held-out test set, where the model struggles to adapt to differences in pre-release preprocessing strategies. A targeted qualitative review of 160 representative scans further reveals key factors contributing to both successful and failed segmentations, including variations in image field of view, temporal enhancement patterns, acquisition era, and artifact prevalence. Overall, these findings underscore the importance of accounting for dataset heterogeneity, domain shift, and standardized preprocessing in the development of robust, clinically deployable breast MRI segmentation models capable of generalizing across institutions and imaging protocols.

Сведения · Details