Abstract

Info

Title:: Integration of prototype generation and contrastive learning for liver tumor segmentation in ultrasound image

Author(s):: ZHANG Congrui1; ZHANG Xukun2; HAN Minghao2; ZHANG Lihua2; WANG Xiaoying3; 1. School of Health Science and Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China 2. Academy for Engineering and Technology, Fudan University, Shanghai 200082, China 3. Department of Liver Surgery, Zhongshan Hospital, Fudan University, Shanghai 200082, China

Keywords:: Keywords: ultrasound image liver tumor segmentation prototype generation contrastive learning Transformer

Abstract:: Abstract: To tackle the challenges posed by low resolution, weak contrast, and positional variations in liver tumor ultrasound images, which affects the diagnostic efficiency and accuracy, a novel method based on prototype generation and contrastive learning is proposed for liver tumor segmentation in ultrasound images. The core of this method is a weighted mask attention Transformer structure which utilizes the probabilities of real categories in the predicted probability distribution to weight image feature vectors, generates class prototypes with category discrimination, and thereby effectively captures key features while enhancing spatial information representation. By combining contrastive loss with Dice cross-entropy loss, the model achieves significant improvements in both category discrimination capability and segmentation accuracy, and overcomes the limitations of traditional models related to insufficient spatial information and intra-class pixel distribution imbalance. Comprehensive evaluations of the proposed method are conducted on a collected dataset of 253 ultrasound images, and the experimental results reveal that the proposed method attains a mean intersection-over-union of 78.44% and a Dice similarity coefficient of 87.41%, validating its superiority in liver tumor segmentation from ultrasound image.