A Comparative Study of Medical Image Segmentation via ResNet-Driven UNet Architecture

Medical image segmentation is a critical component of medical image processing and analysis, aimed at precisely delineating the target region from the background and facilitating the extraction of pertinent semantic characteristics. This research proposes an upgraded U-Net architecture utilizing a ResNet encoder to augment the segmentation accuracy of benign and malignant breast lesions by integrating the deep feature extraction capabilities of the residual network with the multi-scale feature reconstruction advantages of U-Net. The original U-Net encoder is substituted with ResNet18 and ResNet50, employing its multi-stage residual blocks to extract hierarchical features, which are dynamically fused with decoder features via skip connections. Experiments on the Dataset BUSI with GT dataset demonstrate that the ResNet-based U-Net surpasses the original model. Subsequent investigation indicates that ResNet18-U-Net attains the optimal equilibrium between computing efficiency and performance, rendering it particularly appropriate for the rapid screening of malignant lesions. ResNet50-U-Net exhibits the lowest false detection rate due to its precise modeling of benign lesion morphology using deep characteristics.