Comparison of Loss Functions for Fibroglandular Tissue Segmentation in MRI.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Accurate segmentation of breast, fibroglandular tissue (FGT), and implants in MRI is crucial but challenging due to anatomical variability and imaging heterogeneity.
• Loss functions significantly impact segmentation performance in medical imaging tasks.

Purpose of the Study:

• To compare the efficacy of three distinct loss function strategies for segmenting breast, FGT, and implants in breast MRI.
• To evaluate a baseline Dice + cross-entropy loss, Skeleton Recall Loss (SRL), and a novel Volume-Aware Loss (VAL).

Main Methods:

• Utilized 136 pre-contrast T1-weighted breast MRI volumes.
• Implemented a 5-fold cross-validation approach.
• Compared segmentation performance using Dice Similarity Coefficient (DSC) for breast, FGT, and implants.

Main Results:

• The baseline model with Dice + cross-entropy loss achieved the highest overall performance (mean DSC: 0.921 breast, 0.712 FGT, 0.933 implant).
• Skeleton Recall Loss (SRL) and Volume-Aware Loss (VAL) showed comparable performance to the baseline within statistical variance.
• No significant improvement was observed with SRL or VAL over the standard approach in this dataset.

Conclusions:

• The standard Dice + cross-entropy loss function is the current gold-standard for heterogeneous breast MRI segmentation.
• Novel loss functions like SRL and VAL did not demonstrate superior performance in this specific application.
• Further research may be needed to optimize SRL and VAL for complex medical imaging segmentation tasks.