Analysis of two variants of U-net for pulmonary nodule segmentation: attention U-net and Dense-Attention U-net

Abstract

Lung nodules, solid or subsolid lung masses smaller than 3 centimetres, including subtle nodules within complex lung tissue, can go unnoticed by medical professionals due to fatigue or limited expertise. To address this challenge, our study proposes an algorithm for lung region of interest (ROI) computed tomography (CT) image processing based on the Attention U-net architecture and an enhanced variant called Dense-Attention U-net. The Attention U-net incorporates Attention Gates in the decoding path, facilitating the passage of relevant information while reducing irrelevant learning. We evaluate model performance using Dice loss and receiver operating characteristic (ROC) curve analysis. The Dense-Attention U-net enhances the model with dense connectivity in both encoder and decoder sections, ensuring complete layer connections. We used a dataset of 27,190 lung CT images for evaluation. Both U-net variants perform well, with the Dense-Attention U-net outperforming the Attention U-net. The Attention U-net took about eight hours to reach a training loss of 0.13, while the Dense-Attention U-net achieved the same in just half an hour. Notably, the Dense-Attention U-net achieves higher predictive accuracy, with area under the curve (AUC) values of 0.94 and 0.91 for the ROC curves, respectively. Visual results demonstrate excellent segmentation performance for both models. In conclusion, our study introduces and analyses two U-net variants for pulmonary nodule segmentation, emphasizing attention mechanisms and dense connections to enhance feature focus and model efficiency. We acknowledge challenges such as dataset biases and suggest future research directions, including individual nodule labeling and quantification, to enhance diagnostic accuracy.