LungAide: An Effective 3D Transformer V-Net for Lung Nodule Detection

Existing lung nodule detection methods on computed tomography (CT) have the problems of missing small nodules and losing spatial information. This paper proposes LungAide, a novel 3D Transformer V-Net for accurate lung nodule detection in chest CT scans, with a particular focus on small nodules. The proposed architecture integrates the encoder-decoder network and concatenation operation of V-Net with the Transformer blocks featuring a lightweight self-attention mechanism with a Bi-Path downsampling, enabling to preserve spatial information and capture long-range dependencies between lung nodules. To validate the effectiveness of LungAide, experiments are conducted on the LUNA16 dataset, which achieving a Competition Performance Metric of 0.915. Comparative and ablation studies confirm that LungAide effectively reduces false positives while enhancing sensitivity.