Abstract:
Self-supervised learning (SSL) has achieved remarkable progress in medical image segmentation. SSL can effectively learn knowledge from limited labeled data and obtain re...Show MoreMetadata
Abstract:
Self-supervised learning (SSL) has achieved remarkable progress in medical image segmentation. SSL can effectively learn knowledge from limited labeled data and obtain reliable prediction results while reducing the training cost and the workload of annotators. As the training progresses, the cognitive bias of the model will also deepen. The cognitive bias of the model may lead to wrong boundary detection, which reduces the accuracy and quality of segmentation. In this paper, we propose a dual-subnet mutual correction framework. The VNet is adopted as subnet 1. Subnet 2 is a variant of VNet that employs ResNet34 as the encoder and introduces the 3D Squeeze-and-Excitation module (SE-R34VNet) to enhance the feature extraction and fusion. By employing mutual learning and correction between two subnets, this framework effectively reduces the cognitive bias of the model and enhances the segmentation performance of the network. We validate the proposed method on the LA dataset. Experiments show that the Dice and Jaccard reach 89.52% and 81.74% respectively. The 95HD drops 1.32 and ASD drops 0.3. Experimental results show that the proposed method outperforms several state-of-the-art methods in the medical image segmentation.
Published in: 2024 IEEE 7th Advanced Information Technology, Electronic and Automation Control Conference (IAEAC)
Date of Conference: 15-17 March 2024
Date Added to IEEE Xplore: 25 April 2024
ISBN Information: