Prediction of visceral pleural invasion in lung cancer on CT: deep learning model achieves a radiologist-level performance with adaptive sensitivity and specificity to clinical needs

Objectives To develop and validate a preoperative CT-based deep learning model for the prediction of visceral pleural invasion (VPI) in early-stage lung cancer. Methods In this retrospective study, dataset 1 (for training, tuning, and internal validation) included 676 patients with clinical stage IA lung adenocarcinomas resected between 2009 and 2015. Dataset 2 (for temporal validation) included 141 patients with clinical stage I adenocarcinomas resected between 2017 and 2018. A CT-based deep learning model was developed for the prediction of VPI and validated in terms of discrimination and calibration. An observer performance study and a multivariable regression analysis were performed. Results The area under the receiver operating characteristic curve (AUC) of the model was 0.75 (95% CI, 0.67-0.84), which was comparable to those of board-certified radiologists (AUC, 0.73-0.79; all p > 0.05). The model had a higher standardized partial AUC for a specificity range of 90 to 100% than the radiologists (all p < 0.05). The high sensitivity cutoff (0.245) yielded a sensitivity of 93.8% and a specificity of 31.2%, and the high specificity cutoff (0.448) resulted in a sensitivity of 47.9% and a specificity of 86.0%. Two of the three radiologists provided highly sensitive (93.8% and 97.9%) but not specific (48.4% and 40.9%) diagnoses. The model showed good calibration (p > 0.05), and its output was an independent predictor for VPI (adjusted odds ratio, 1.07; 95% CI, 1.03-1.11; p < 0.001). Conclusions The deep learning model demonstrated a radiologist-level performance. The model could achieve either highly sensitive or highly specific diagnoses depending on clinical needs.

Automated summary

The study developed and validated a deep learning model based on preoperative CT for predicting visceral pleural invasion in early-stage lung cancer. The model demonstrated comparable performance to board-certified radiologists and could provide highly sensitive or highly specific diagnoses based on clinical needs.