LAUSR

Background: It is significant to evaluate the degree of invasiveness of lung adenocarcinomas (LACs) appearing as subsolid nodules (SSNs) in CT before making clinical management decisions. The role of automated deep learning (DL) method in assisting doctor to classify different levels of malignancy remains unclear. Aims: To explore the diagnostic performance and clinical utility of a 3D DL method in malignancy assessment of SSNs. Methods: CT data of patients with SSNs between 2013.1-2015.12 was reviewed and those pathologically diagnosed of LACs were collected, which were then randomly separated into development (85%) and test sets (15%). A 3D DL model was trained using the development set. The diagnostic performance was evaluated by comparing with those of three doctors and the validated Brock model on the test set. And the clinical utility was further evaluated in a prospective database. Results: A total of 1589 SSNs from 1471 patients were included, of which 66.2% were female and the mean age was 56 years (range 23-84). In the differentiation of invasive adenocarcinoma, the automated model achieved a similar AUC of 0.91 (95%CI, 0.85-1) with that of doctors (0.91, [95%CI 0.90-0.94]). Brock model achieved the lowest AUC of 0.82 (95%CI, 0.77-0.86). In multi-class evaluation, AUCs of doctors were significantly improved with the help of model in both four-class and three-class (when merging AAH and AIS together) to 0.87 (95%CI, 0.84-0.90) and 0.87 (95%CI, 0.83-0.90), respectively. In prospective validation, the model reached an equivalent AUC with doctors. Conclusions: DL method achieved similar diagnostic performances to doctors, and it will potentially improve accuracy and efficiency in SSNs evaluation.