DSENet: Double three-dimensional squeeze-and-excitation network for pulmonary nodule detection

The early detection and diagnosis of lung cancer is critical for improving the survival rate of patients. Deep learning networks are being increasingly applied in the field of medically assisted diagnosis. Although a nodule detection method based on deep learning has achieved good performance, building a nodular detection network with good generalization performance is still challenging owing to the unbalanced positive and negative samples. To overcome this problem, a single-stage double three-dimensional (3D) squeeze-and-excitation (SE) network was proposed for pulmonary nodule detection without using false-positive reduction. The double 3D SE network is extended from a two-dimensional (2D) SE network and explicitly learns the relationship between channels and the interdependence between the depths in each channel. Usually, only a small number of group truth 3D pulmonary nodules are present in a computed tomography (CT) image; thus, another nodule was added to each CT image, and the additional nodules originate from the other CT images in the training dataset. We validated our method using the public lung nodule dataset LUNA16 and achieved a sensitivity of 97.3%. We achieved better results compared to the best single-stage method currently available.