A lightweight convolutional neural network based on U shape structure and attention mechanism for anterior mediastinum segmentation

To automatically detect anterior mediastinum lesions (AMLs) in the anterior mediastinum (AM), an automatic segmentation model designed explicitly for AM regions in chest computed tomography (CT) scans is required. Due to the low prevalence of AML, reviewing large CT datasets retrospectively is time-consuming. Developing an artificial intelligence (AI) model to identify the AM region can help radiologists manage workloads and improve diagnostic accuracy. In this paper, we introduce a U-shaped network architecture with two attention mechanisms to maintain long-range dependencies and enhance localization. We propose a parallel multi-head self-attention (MHSA) module called wide-MHSA (W-MHSA), along with a dilated depth-wise parallel path connection (DDWPP) to support upsampling stages. Additionally, an expanding convolution block is combined with W-MHSA in the encoder to ensure a lightweight design. Our proposed model demonstrates superior segmentation performance compared to state-of-the-art networks, showing strong potential for clinical application in AM lesion detection workflows.