A Medical Image Semantics Segmentation Method Based on Image Pre-processing and Image Transformer

School of Mathematics and Statistics, University of Melbourne, Melbourne, 3052, Australia

^* Corresponding author: zhaopengl@student.unimelb.edu.au

Abstract

Semantics segmentation is a task aiming at classifying each pixel of an image into a category. Because of its ability to comprehend and interpret the context of images, it plays a vital role in the medical area. Existing models in this field pay more attention to editing U-Net’s structure and convolutional layers. In this paper, an image pre-processing technique to enrich the potential of images and an image transformer involved in network TrUNet are proposed to tackle these issues. In particular, the histogram equalisation method was introduced to improve the quality of each image, and image augmentation methods including flipping and rotation were used to increase the size of the image dataset. Then TrUNet cooperates vision transformer to U-Net to improve the model’s ability to recognize global context. Extensive experiments show that this method outperforms U-Net on the DRIVE dataset. Its validation loss curve is smoother and decreases more decently. Its accuracy increases more significantly in early epochs and is higher than that of U-Net. It has higher precision and F1-score compared with U-Net.