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

Biomedical Image Segmentation applications have witnessed mushroom growth in the last two decades. Current state-of-the-art approaches face challenges when dealing with instance imbalances in datasets. Various functions, such as Blob Loss, Lesion-wise Loss, and Dice Loss limitations, were addressed by Instance-wise loss and Center-of-Instance loss (ICI). ICI is the result of Instance loss, and the center of instance loss suffers from highly unregulated labels and outputs, resulting in low accuracy of aforementioned loss functions. We introduce a novel dual-coefficient regularization approach for loss functions that modifies both predicted outputs and labels before loss computation. This addresses instance imbalance more effectively than previous pixel-level or class-level weighting strategies. The proposed approach resulted in the enhancement of existing loss functions: (1) RIW (regularized instance-wise loss), (2) RCI (regularized center of instance loss), and (3) RPW (regularized pixel-wise loss). The simulation experiments on the ATLAS R2.0 (MICCAI, 2022) and BraTS'20 (MICCAI, 2020) datasets validated our approach in comparison with the state-of-the-art loss functions resulting in significant improvements in RIW (up to 69.16%), RCI ( up to 16.58%), RPW (67.82%), subsequently decreased false detection rate up to (97.78%), and number of missed instances.