ABSTRACT
Objective: To investigate the feasibility of deep convolutional neural network (CNN) to classify the pathological type of sub-soild pulmonary adenocarcinoma of ≤3 cm based on CT images, and to visualize the medical imaging features derived from the activation area of CNN. Methods: A total of 200 sub-solid lung nodules, which were confirmed as adenocarcinoma by immunohistochemical staining, were classified as preinvasive lesions (including atypical adenomatous hyperplasia and adenocarcinoma in situ), microinvasive adenocarcinoma and invasive adenocarcinoma, in which 160 (80%) were used to train the inception v3 CNN architecture, and the other 40 (20%) were used to test the model and visualize the activation area. The characteristics of the activated area were defined as 14 CT signs. Results: The CNN yielded an accuracy of 87.5% to classify three categories of lung nodules. The visualization study found that the CNN activation area mainly focused on the non-solid component (43.0%) and smooth margin (20.2%) of the preinvasive lesions, on the spiculated margin (18.3%) of the microinvasive adenocarcinoma, and on the solid component (18.9%) and the spiculated margin (14.1%) of the invasive adenocarcinoma. Conclusion: CNN can classify the pathological type of lung adenocarcinoma based on CT images. The visualization of activation area of CNN indicates the medical imaging characteristics of diagnosis.