A survey of U-Net variant network for MRI brain tumor segmentation

Accurate segmentation of tumors and lesions in biomedical images is essential for assisting clinicians in assessing disease severity and developing effective treatment plans. Among various segmentation methods, U-Net has become the most widely adopted architecture due to its strong performance and suitability for medical imaging tasks. This paper presents a comprehensive review of U-Net and its recent variants. First, the structural principles and core algorithms of the original U-Net architecture are introduced. Then, recent U-Net-based models are categorized into three major groups: (1) attention mechanism, (2) transformer-based mechanism, and (3) skip connection enhanced mechanism. The structural differences, advantages, limitations, and applicable scenarios of each category are analyzed in depth. Furthermore, the paper assesses the latest advancements in biomedical image segmentation and identifies the current limitations of U-Net-based approaches. Based on these findings, a new design scheme is proposed to guide future improvements in U-Net architecture. This work aims to provide researchers with a clear understanding of existing U-Net variants and offer novel insights for advancing medical image segmentation.