Evaluating Medical Image Segmentation Models: A Critical Review of Current Methods and Future Directions

Since the emergence of the U-Net /cite{3,7}, it has exhibited significant accomplishments in the field of medical image segmentation. However, given the increasing intricacy and novelty of clinical application scenarios and tasks, the conventional U-Net falls short in meeting the requirements. Consequently, numerous models, such as Attention U-Net /cite{6}, nnU-Net /cite{5}, and Transunet /cite{8}, have been developed and have achieved state-of-the-art outcomes in various clinical medical image segmentation tasks and challenges /cite{12,14,15}. These advancements have greatly facilitated the progression of medical image segmentation. /n/nDespite the rapid evolution of medical image segmentation models, there remains a scarcity of evaluation methods specifically tailored to particular clinical application scenarios. Moreover, the existing methods suffer from deficiencies in terms of comprehensiveness, complexity, and consistency. This paper argues that the development of robust and comprehensive evaluation frameworks is crucial for advancing the field of medical image segmentation. It proposes that future research should focus on developing evaluation methods that are sensitive to the unique challenges and requirements of different clinical applications. This will ensure that the models developed are not only accurate but also clinically relevant and reliable.