Identifying the important components in medical and brain image segmentation framework

Abstract

The U-Net model has been the current inspiration for many of the SOTA medical image segmentation models. Many models, such as U-Net++ and Attention U-Net are variations of the U-Net architecture. The nnU-Net model improves upon the base U-Net architecture by creating an AutoML way of creating a specific configuration for each specific dataset. Though this results in an improvement to the base U-Net model, this can be further refined by conducting ablation tests to find out the most important components of this model. In doing so, we can improve on all models that make use of U-Net, as they would also be able to make use of nnU-Net. In this study, ablation tests were conducted on the Medical Segmentation Decathalon dataset. Experimental findings have shown that increasing the batch size and changing the target spacing can improve performance.