Breast Tumor Segmentation in Dynamic Contrast-Enhanced Magnetic Resonance Images via Multi-Staged Training and Deep Ensembling of a Large Kernel MedNeXt

Toufiq Musah

Keywords: Breast Tumors, Segmentation, Large Kernel, MedNeXt

TL;DR: We segment breast tumors in DCE-MRI using a multi-stage trained MedNeXt, using kernel expansion to enlarge receptive fields and employing optimization-based loss ensembling for improved performance.

Abstract: Accurate segmentation of breast tumors in dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) is essential for disease assessment and treatment planning but remains challenging due to limited annotated data and high variability in enhancement patterns. In this work, we present a MedNeXt-based segmentation framework trained through a multi-staged process designed to improve receptive field coverage and model robustness. We first train a conventional MedNeXt model with 3 x 3 x 3 kernels and then expand its convolutional receptive fields using a trilinear interpolation procedure called UpKern, forming a large-kernel variant with 5 x 5 x 5 kernels. The large-kernel models are fine-tuned and combined through deep ensembling, including optimization with a hybrid loss that balances Dice–cross-entropy and focal loss to address small lesion errors and class imbalance. Experiments on a multicenter breast DCE-MRI dataset of over 1,500 cases demonstrate consistent improvements in Dice score and normalized Hausdorff distance compared to both the baseline MedNeXt and a five-fold nnU-Net ensemble, despite using fewer parameters per model. The results indicate that receptive field expansion through UpKern, and loss optimization-based ensembling can enhance tumor segmentation performance while maintaining computational efficiency.

Submission Number: 32