The Combination of Attention Sub-convnet and Triplet Loss for Pulmonary Nodule Detection in CT Images

This paper proposed models based on CNN to detect lung cancer tumors in CT images. More details, three models combined multiple Convolutional Attention Networks were generated: (1) ATT (Attention-Triplet-Triplet) used triple loss in training and testing; (2) ASS (Attention–Softmax–Softmax) used Softmax loss in training and testing; (3) AST (Attention–Softmax–Triplet), AST (Attention-Softmax–Triplet) used ASS as a pre-trained model in training and triplet loss in testing. Theoretical and empirical analyses were discussed to demonstrate the efficacy of the AST model in comparison with ATT and ASS. The feasibility of the AST model is also confirmed when compared to other methods on the same dataset (AST obtained has a specificity of 98.9%).