A novel regularization approach for loss functions to reduce instance imbalance in biomedical image segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Biomedical image segmentation has grown significantly, but current methods struggle with imbalanced datasets.
• Existing loss functions like Blob Loss, Lesion-wise Loss, and Dice Loss have limitations, particularly with instance imbalance.
• Instance-wise loss and Center-of-Instance loss (ICI) were developed but suffer from unregulated labels and outputs, leading to low accuracy.

Purpose of the Study:

• To introduce a novel dual-coefficient regularization approach for loss functions in biomedical image segmentation.
• To address the challenge of instance imbalance more effectively than existing pixel-level or class-level weighting strategies.
• To enhance the performance of existing loss functions through regularization.

Main Methods:

• A novel dual-coefficient regularization approach was developed, modifying predicted outputs and labels before loss computation.
• This approach was applied to create enhanced loss functions: regularized instance-wise loss (RIW), regularized center of instance loss (RCI), and regularized pixel-wise loss (RPW).
• Experiments were conducted on the ATLAS R2.0 and BraTS'20 datasets.

Main Results:

• The proposed approach significantly improved existing loss functions: RIW (up to 69.16%), RCI (up to 16.58%), and RPW (67.82%).
• A substantial decrease in the false detection rate was observed, up to 97.78%.
• The number of missed instances was also reduced, demonstrating enhanced segmentation accuracy.

Conclusions:

• The dual-coefficient regularization approach effectively addresses instance imbalance in biomedical image segmentation.
• The enhanced loss functions (RIW, RCI, RPW) demonstrate superior performance compared to state-of-the-art methods.
• This method offers a promising solution for improving the accuracy and reliability of biomedical image segmentation.