Boosting UNet Performance Via VGG-Based Encoder for Medical Image Segmentation

Aiming at the problems of blurred edges and missed detection of small targets caused by the insufficient feature extraction ability of shallow encoders in the UNet model for medical image segmentation tasks, this study proposes an improvement method of UNet based on the enhancement of VGG encoders. By replacing the original encoder of UNet with pretrained VGG16 and VGG19 networks, its deep symmetric structure and continuous small convolutional kernel stacking are utilized to enhance the multi-scale feature expression capability, while retaining the jump connections to achieve cross-layer feature fusion. Experiments on the breast ultrasound image dataset BUSI-with-GT show that the improved VGG19-UNet improves over the original UNet regarding the Dice coefficient, mIoU, and mPA. This study verifies the effectiveness of migration learning in medical image segmentation and provides new ideas for designing segmentation models with both lightweight and high accuracy.