Predicting Future Lung Cancer Risk from a Single Low-Dose CT Using Deep Learning

Take-Away Points

• ■ Major Focus: Developing a deep learning model, named Sybil, to predict lung cancer risk over 6 years based on a single chest low-dose CT (LDCT).

• ■ Key Results: Sybil predicted the risk of lung cancer with optimal performance, which was further confirmed on two large independent external data sets.

• ■ Impact: If further validated, Sybil may be a fast and easily implementable tool to improve and personalize patient management within lung cancer screening.

Although LDCT-based lung cancer screening (LCS) has resulted in a substantial reduction in lung cancer mortality, several aspects, including the definition of individual risk, must be improved to avoid unnecessary invasive procedures and to reduce long-term radiation exposure and economic costs.

In their study, Mikhael and Wohlwend et al validated a deep learning algorithm to predict lung cancer risk from a single chest LDCT. The authors retrospectively analyzed a single scan from the National Lung Screening Trial (NLST) to develop a model for predicting lung cancer risk over 6 years and validated the algorithm in two independent LCS-LDCT data sets from Massachusetts General Hospital and Chang Gung Memorial Hospital (Taiwan). In the NLST data set, Sybil achieved area under the receiver operating characteristic curve (AUC) values of 0.92 (95% CI: 0.88, 0.95) at 1 year and 0.86 (95% CI: 0.82, 0.90) at 2 years and a C-index of 0.75 (95% CI: 0.72, 0.78) over 6 years. Performances were comparable in the two external data sets and remained consistent across sex, age, and smoking history NLST subgroups.

To differentiate between the Sybil's ability to predict future cancer risk and detect existing cancer, the authors examined model performance after removing LDCT scans showing lung nodules subsequently proven as cancer. Sybil's performance decreased but remained predictive, with AUCs of 0.81 (95% CI: 0.74, 0.86) at 2 years and 0.69 (95% CI: 0.63, 0.74) at 6 years.

Although further studies are needed to assess the clinical applications of Sybil, this model has potential to personalize LCS according to the individual risk of patients and would be relatively easy to implement, requiring only LDCT images and no manual segmentation. Before proceeding with prospective assessment of Sybil's performance, this deep learning model requires further validation to assess its generalizability in a wider and more diverse population and identify any potential shortcomings.

Highlighted Article

• Mikhael PG, Wohlwend J, Yala A, et al. Sybil: A Validated Deep Learning Model to Predict Future Lung Cancer Risk From a Single Low-Dose Chest Computed Tomography. J Clin Oncol 2023; 41(12):2191–2200. doi: https://doi.org/10.1200/JCO.22.01345.