Deep Learning-driven Segmentation of Breast Cancer MR Imaging Using a Hybrid ResNet-Transformer Architecture and Cross-dataset Validation.

Authors

Abstract

To develop and cross-dataset validate a hybrid ResNet50-Transformer U-Net for automated segmentation of breast cancer (BC) lesions on 2D dynamic contrast-enhanced MRI (DCE-MRI). A total of 20434 BC MRI images from the BC MRI Segmentation Benchmark (BC-MRI-SEG), comprising 4 independent public datasets, were included. A strict patient-level data splitting strategy was applied to avoid data leakage. The proposed model was based on a U-Net architecture integrating a ResNet50 encoder for local feature extraction and a bottleneck Transformer module for global contextual modeling. Model performance was evaluated using a leave-one-dataset-out cross-dataset external validation strategy. Segmentation performance was assessed using the Dice similarity coefficient (DSC), intersection over union (IoU), 95% Hausdorff distance (HD95), sensitivity, and precision. The proposed hybrid model achieved stable segmentation performance across the evaluated datasets. In leave-one-dataset-out external validation, patient-level DSC ranged from 0.75 ± 0.13 to 0.79 ± 0.10, while IoU ranged from 0.61 ± 0.15 to 0.66 ± 0.12. In the fixed held-out comparison cohort, the proposed model achieved a DSC of 0.82 ± 0.08 and an IoU of 0.70 ± 0.10 and outperformed the evaluated baseline models (Holm-adjusted P ≤ 0.021). Qualitative analysis showed generally close agreement between the predicted masks and reference annotations across lesions with different morphologic appearances. The proposed ResNet-Transformer-based U-Net framework provided a robust approach for automated BC MRI segmentation across heterogeneous public datasets. These findings suggest that combining local feature extraction with global contextual modeling may be useful for computer-assisted breast MRI analysis and related clinical workflows.

Tags

Topics