Applications of deep learning in medical image segmentation: an evaluation of U-Net architecture and its enhancements

In modern healthcare, medical image segmentation (MIS) technology is an important reference for clinical diagnosis, especially for illness diagnosis, treatment planning, and surgical navigation. Because of its exceptional performance and adaptability, the U-Net architecture has emerged as a leading method in the medical picture segmentation area with the ongoing advancements in deep learning technology. The unique symmetric U-shaped architecture and skip connections of the U-Net structure are noteworthy because they allow for the accurate recording of both local and global characteristics in medical imaging activities. Nevertheless, U-Net faces certain limitations in handling long-term dependencies and extracting global features. This research encompasses a series of experiments employing the U-Net framework to explore and assess its efficacy, alongside several modified versions, in the context of Medical Image Segmentation (MIS). The primary objective is to enhance the performance of MIS through comprehensive evaluation. The findings demonstrate that the U-Net model achieves high-precision segmentation when optimal training procedures and data augmentation techniques are applied. These results are based on an examination using a dataset of drosophila larva ventral nerve cord serial sections, which provides a robust benchmark for segmentation accuracy. Notably, the performance is particularly impressive at lower learning rates.