Enhancing CT image segmentation accuracy through ensemble loss function optimization.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computational Imaging

Background:

• Deep neural network training for CT medical image segmentation is sensitive to loss function choice.
• Individual loss functions (Cross Entropy, Dice, Boundary, TopK) have limitations and biases.

Purpose of the Study:

• To enhance segmentation accuracy by optimizing ensemble loss functions.
• To mitigate biases and limitations inherent in single loss functions and their linear combinations.

Main Methods:

• Evaluated integrated loss functions (Cross Entropy, Dice, Boundary, TopK) via linear combination and model-level ensembles.
• Utilized Attention U-Net (AttUNet) and SwinUNETR architectures on institutional and public CT datasets.
• Employed static averaging and learnable dynamic weighting for ensemble strategies.

Main Results:

• Ensemble methods increased Dice Similarity Coefficient (DSC) scores by 2%-7% compared to non-ensemble approaches.
• Achieved significant reductions in Hausdorff Distance (HD) and Average Surface Distance (ASD), e.g., 19.1% HD reduction for rectum segmentation.
• Learnable ensembles with optimized weights produced finer segmentation details and outperformed static ensembles.

Conclusions:

• Ensemble models with optimized weights effectively improve medical image segmentation accuracy.
• This approach shows potential for advancing automated medical image analysis applications.