2041 - A 3D CNN Model Using Deep Learning Based on Chest CT for Predicting Disease-Free Survival of Patients with Early Stage Lung Cancer Receiving Surgery and SBRT

Y. Fu¹, R. Hou¹, X. Fu¹, L. Qian², W. Feng¹, Q. Zhang¹, Z. Ding², W. Yu¹, X. Cai¹, and J. Liu¹; ¹Department of Radiation Oncology, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China, ²Department of Oncological Surgery, Shanghai Lung Cancer Center, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China

Purpose/Objective(s): Although surgery and stereotactic body radiation therapy (SBRT) are curative treatment strategies for patients with early stage non-small cell lung cancer (NSCLC), tumor recurrence is not uncommon after both treatments leading to poor survival outcomes, suggesting the necessity of subsequent medical treatments for specific patients with poor prognosis. Therefore, we tried to construct a 3D CNN model based on the pretreatment CT images to predict prognosis of clinical stage I NSCLC patients undergoing surgery, and then applied it in an SBRT cohort to help improve their prognosis prediction performance which might be poorer due to the lack of definitive pathological data, thus benefiting more precise individualized treatment. Materials/

Methods: Our review retrospectively included a surgical cohort made up of clinical stage I NSCLC patients with surgical resection between 2015 and 2017 and an SBRT cohort with clinically-diagnosed early stage lung cancer from 2014 to 2020 in our institution. A 3D CNN model was constructed from preoperative chest CT images of the surgical cohort to achieve accurate prediction risk scores of DFS and stratify different risk patients according to the best cutoff value, which was then test in the SBRT cohort using their pretreatment CT examinations. Kaplan-Meier method and log-rank test were used to estimate survival differences of different risk groups. The hazard ratios were calculated with the risk score and clinical factors by using the multivariate Cox regression. The concordance index (C-index) was analyzed to evaluate discrimination of our model in both cohorts.
Results: A total of 2489 and 153 patients who met the inclusion criteria were included in this study to form the surgical and SBRT cohort, respectively. Based on CT images, a DFS predictive model built by deep learning capable of performing accurate prognostic risk stratification for surgical patients with clinical stage I NSCLC was developed with a C-index significantly better than the clinical model (0.857 vs 0.747, 0.855 vs 0.750 and 0.847 vs 0.764 in the training, validation and testing set, P<0.001). When applied in the SBRT cohort with clinical diagnosis of early-stage lung cancer, external verification was carried out and the K-M curve could significantly distinguish high- and low-risk groups (P<0.0001). Besides, the multivariate Cox regression analysis showed that our model output was one of the independent predictors of DFS in both cohorts (P<0.001).
Conclusion: An end-to-end 3D CNN model for prognosis prediction of patients with early stage lung cancer was successfully constructed for different risk stratification of recurrence both in surgical and SBRT patients, and was demonstrated as a factor independent from other clinical features for DFS prediction.