Anterior mediastinal nodular lesion segmentation from chest computed tomography imaging using UNet based neural network with attention mechanisms

Automated detection of anterior mediastinal nodular lesions (AMLs) has significance for clinical usage as it is challenging for radiologists to accurately identify AMLs from chest computed tomography (CT) imaging due to various factors, including poor resolution, variations in intensity and the similarity of the AMLs to other tissues. To assist radiologists in AML detection from chest CT imaging, a UNet-based computer-aided detection (CADe) system is proposed to segment AMLs from slice images of the chest CT scans. The proposed network adopts a modified UNet architecture. To guide the proposed network to selectively focus on AMLs and potentially disregard others in the image, different attention mechanisms are utilized in the proposed network, including the self-attention mechanism and the convolutional block attention module (CBAM). The proposed network was trained and evaluated on 180 chest CT scans which consist of 180 AMLs. 90 AMLs were identified as thymic cysts, and 90 AMLs were diagnosed as thymoma. The proposed network achieved an average dice similarity coefficient (DSC) of 93.23 with 5-fold cross-validation, for which the mean Intersection over Union (IoU), sensitivity and specificity were 90.29, 93.98 and 95.68 respectively. Our method demonstrated an improved segmentation performance over state-of-the-art segmentation networks, including UNet, ResUNet, TransUNet and UNet++. The proposed network employing attention mechanisms exhibited a promising result for segmenting AMLs from chest CT imaging and could be used to automate the AML detection process for achieving improved diagnostic reliability.