Adaptive composite loss for volumetric whole heart segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate medical image segmentation is crucial for diagnosis and treatment planning.
• Existing loss functions often struggle to balance regional overlap and precise boundary delineation.
• Developing effective loss functions is key to advancing automated medical image analysis.

Purpose of the Study:

• To evaluate composite loss functions for medical image segmentation.
• To compare fixed and adaptive weighting schemes for combining binary cross-entropy (BCE) and boundary-based loss terms.
• To assess the impact of model architecture (U-Net vs. SwinUNETR) on the effectiveness of different loss functions.

Main Methods:

• Utilized the MM-WHS dataset for segmentation tasks.
• Implemented U-Net and SwinUNETR architectures.
• Evaluated composite losses: Binary Cross-Entropy (BCE) with a Boundary-based term (BoundaryDoU).
• Compared fixed weighting ratios against adaptive weighting schemes (e.g., SoftAdapt).

Main Results:

• For U-Net, adaptive weighting with a small boundary contribution (90/10 BCE + BoundaryDoU) achieved the highest Dice score.
• SwinUNETR demonstrated peak performance with a fixed weighting ratio (70% BCE + 10% boundary).
• The study observed that boundary-based loss terms enhance segmentation accuracy, but gains are architecture-dependent.

Conclusions:

• Composite loss functions integrating boundary information improve segmentation accuracy in medical imaging.
• Convolutional neural networks like U-Net benefit from adaptive loss weighting schemes.
• Transformer-based models like SwinUNETR may require different optimization strategies for boundary-aware segmentation.