ResUNet Model Enhanced with Multiple Attention Mechanisms for Effective Pulmonary Nodule Segmentation in CT Images

Abstract

This research proposes the Position combined Channel attention module - Residual Unet model (PCAM-ResUnet), an enhanced ResUnet++, to improve CT lung nodule segmentation. In the paper, the Squeeze-and-Excitation Block is replaced by Channel Attention module (CAM) and Position Attention module (PAM) respectively. More importantly, these two modules are combined to create the Position combined Channel attention module (PCAM), a new breakthrough in our model structure. Through a multi-stage training process, PCAM-ResUnet was evaluated on a test dataset comprising 2000 pulmonary nodule samples. The PCAM variant demonstrated outstanding performance, achieving an average Dice Similarity Coefficient (DSC) of 85.96%.  It achieved 'Excellent' segmentation results (cases with a DSC ≥ 80%) in 82.80% of cases, while reducing the 'Needs Improvement' level (DSC < 40%) to 1.85%. The obtained results emphasize the effectiveness of PCAM-ResUnet, affirming its superiority and showcasing its considerable potential for widespread clinical applications in the medical field.