3DMNet: A 3D Multi-Attention Network for Pulmonary Nodule Detection in 3D CT Scans

Artificial intelligence (AI)-assisted diagnosis has become an urgent and active research topic in modern healthcare. The automatic detection of pulmonary nodules through three-dimensional computed tomography (CT) not only significantly reduces the workload of primary care physicians but also aids in the early screening of lung cancer, providing patients with critical additional treatment time. However, existing methods primarily focus on single-slice nodule detection and fail to fully leverage the 3D information embedded within the CT volume, ultimately leading to suboptimal detection accuracy. To overcome this limitation, we propose 3DMNet, a 3D multi-attention convolutional neural network that effectively utilizes the 3D information embedded in multi-dimensional CT data. Specifically, to simulate the clinical diagnostic process of radiologists identifying nodule locations across consecutive CT slices, we develop an innovative 3D spatial attention mechanism. This mechanism enhances feature representations by applying projection transformations across multiple dimensions of the feature map, followed by rigorous attention calculations, thereby capturing intra- and inter-slice spatial information. Furthermore, by integrating data-driven global attention and channel attention modules, we achieve a more refined capture of global spatial dependencies and inter-channel relationships. Extensive comparative analyses on a widely used dataset show that our method achieves an average FROC of 91.33%, demonstrating that the proposed framework outperforms mainstream methods.