Diagnostic Segmentation Based on Kidney Medical Image

Abstract

Lesion segmentation of medical images is an important component of smart medicine. The development of deep learning technology is followed by rapid advancement in lesion segmentation technology of medical images. Though the present segmentation technology can retain spatial features, insufficient spatial features are retained with low segmentation accuracy. Our proposed PST-UNet model combines transformer with U-shaped structure and better infuses encoder’s multiscale features by using convolution fusion module. PST-UNet model adopts two types of block Swin transform at encoder and decoder ends, respectively. Renal lesion data tend to present a normal distribution. Therefore, to preserve more spatial features and enhance the precision of renal lesion segmentation, Swin transformer block and full (Gaussian error linear unit) activation function are introduced at the encoder end. Similarly, at the decoder end, Swin transformer block, full GELU activation function, upsampling, and jumper wires from the convolution fusion module are also introduced.