Abstract

Info

Title:: Lung nodule segmentation algorithm based on full-scale channel feature aggregation coding and decoding network

Author(s):: XIE Shaopeng; WANG Mingquan; GENG Yujie; HUANG Xinyue; SHANG Ran; School of Information and Communication Engineering, North University of China, Taiyuan 030051, China

Keywords:: Keywords: lung nodule segmentation full-scale skip connection binomial constraint loss function dilated convolution

Abstract:: Abstract: To address the difficulty in accurately detecting pulmonary nodules of different properties, a full-scale channel feature aggregation encoding and decoding network (FCFA-Net) is employed to assist experienced physicians in diagnosis. The network which consists of SMC, full-scale feature aggregator, autocorrelation feature enhancer, channel feature hierarchy extraction decoder and binomial constraint loss function can fully extract shallow and deep features from CT images for realizing the segmentation of pulmonary nodules of different sizes and shapes. Compared with UNet, UNet++ and TransUnet, FCFA-Net increases the accuracy by 9.96%, 7.84% and 3.75%, recall rate by 5.50%, 2.96% and 1.37%, mean intersection over union by 11.35%, 7.16% and 4.18%, F1 score by 8.07%, 5.87% and 3.10%, respectively. Additionally, ablation experiment results demonstrate that each structure is effective and can achieve the best result within the acceptable parameter range.