Learn from Each Other: Comparison and Fusion for Medical Segmentation Loss

In recent years, medical image segmentation has been attracting intensive attention from both the industries and academia. Various loss functions have been employed for Medical Image Segmentation toward methods of deep learning. However, little research to date has focused on the summary and analysis of the effects of different loss functions on medical image segmentation. Some existing work focused on the performance of different loss functions did not address the fusion of different types of loss functions. Moreover, some experiments that compare the performance of different loss functions are often limited to brains, full-body regions, etc. There is a lack of study on the segmentation subjects of small volume and large numbers like nucleus for various loss functions. In this paper, we first provide a comprehensive survey on recent progress in Medical Image Segmentation for different loss functions. Then, we compare and analyze the performance of nucleus image segmentation tasks using fusion loss in which a new loss function is proposed. Furthermore, we explain the reason and mechanism underlying the better performance of fusion loss function for nucleus segmentation. Finally, we give suggestions for choosing loss function in medical image segmentation. We believe the analysis on loss function may benefit the industries and academia in Medical Image Segmentation and continuous optimization of the loss function.