Comparison of Loss Functions for Fibroglandular Tissue Segmentation in MRI.

Due to anatomical variability and imaging heterogeneity, the choice of loss function for breast, fibroglandular tissue (FGT), and implant segmentation in magnetic resonance imaging (MRI) is a challenging task. This study compares three loss function strategies for image segmentation: a baseline model using the standard Dice + cross-entropy loss, the Skeleton Recall Loss (SRL) targeting fine structural details, and a novel Volume-Aware Loss (VAL) penalizing implausibly small or false-positive implant predictions. A total of 136 pre-contrast T1-weighted breast MRI volumes were used for 5-fold cross-validation. The baseline achieved the highest overall performance (mean DSC: 0.921 breast, 0.712 FGT, 0.933 implant), while SRL and VAL were comparable to the baseline within the statistical variance of the dataset. Thus, we deduced that the standard Dice + cross-entropy loss remains the gold-standard for heterogeneous breast MRI segmentation.