Integrating Deep Learning of Low-Dose CT Imaging With Clinical Data for Lung Cancer Risk Prediction

Low-dose CT (LDCT) imaging screening reduces lung cancer mortality, the leading cause of cancer deaths globally. Segmentation-free deep learning (DL) models such as Sybil can improve screening efficiency but require extensive validation and possible improvement.

Can the integration of DL based on LDCT scans and clinical data improve lung cancer risk prediction?

Retrospective cohort data from 4 different screening programs, 1 used for model training and 3 used for external validation. Data were collected between 2002 and 2021. The median follow-up period was 7 years. All participants had a history of either current or former smoking, with at least 10 pack-years of smoking or who smoked over 20 years. The area under the receiver operating characteristic curve (AUC) was calculated for lung cancer risk within 1 to 6 years, stratified by pulmonary nodule presence and size. Key clinical and epidemiologic factors were evaluated for their added predictive value.

This analysis used 52,482 LDCT scan series from 22,469 participants. Sybil's AUC ranged from 0.93 in year 1 and reduced to 0.79 in year 6 in the independent cohorts. The predictive performance was suboptimal in the absence of documented nodules (AUC, 0.64) and for small nodules (AUC, 0.61) in year 6. Our new model, Sybil-Epi, trained with baseline scans, achieved higher predictive performance (AUC, 0.83; 95% CI, 0.81-0.85) compared with Sybil (AUC, 0.80; 95% CI, 0.78-0.82) in year 6. The difference is most notable when nodules are absent. Sybil-Epi's AUC was 0.76 (95% CI, 0.70-0.82) and Sybil's AUC was 0.64 (95% CI, 0.57-0.70).

Our results show that Sybil performs better for short-term lung cancer risk, but the predictive accuracy was suboptimal when nodules were absent. Our integrated Sybil-Epi model with DL and clinical and epidemiologic factors significantly improved model predictive performance.