Deep Learning for Fully Automated Prediction of Overall Survival in Patients Undergoing Resection for Pancreatic Cancer

Background: Exploiting prognostic biomarkers for guiding neoadjuvant and adjuvant treatment decisions may potentially improve outcomes in patients with resectable pancreatic cancer. To this end, we develop an objective and robust imaging biomarker for fully automated prediction of overall survival (OS) of pancreatic cancer by directly analyzing multiphase contrast-enhanced CT (CECT) using deep learning.Methods: This retrospective study included 1516 patients with resected pancreatic ductal adenocarcinoma (PDAC) from five centers located in China. The discovery cohort (n=763), which included preoperative multiphase CECT scans and OS data from two centers, was used to construct a fully-automated prognostic biomarker – DeepCT-PDAC – by training a holistic convolutional neural network for volumetric segmentation of PDAC and pancreatic anatomies and four subsequent networks for OS prediction. The marker was independently tested using internal (n=574) and external validation cohorts (n=179) to evaluate its performance, robustness, and clinical usefulness.Findings: Preoperatively, DeepCT-PDAC was the strongest predictor of OS in both internal and external validation cohorts (hazard ratio [HR] 2·03, 95% CI 1·50–2·75, p<0·0001; HR 2·47, 1·35–4·53, p=0·0034) in a multivariable analysis including age, CT tumor size, tumor location, and CA 19-9. Postoperatively, DeepCT-PDAC remained significant in both cohorts (HR 2·49, 95% CI 1·89–3·28, p<0·0001; HR 2·15, 1·14–4·05, p=0·018) after adjustment for resection margin, pT stage, pN stage, tumor differentiation, perineural invasion, pathological tumor size, and treatment. For margin-negative patients, adjuvant radiotherapy was associated with improved OS in the subgroup with DeepCT-PDAC low risk (HR 0·35, 95% CI 0·19–0·64, p=0·00036), but did not affect OS in the subgroup with high risk.Interpretation: Deep learning-derived CT imaging biomarker enabled objective and unbiased prediction of OS for resectable PDAC both pre- and postoperatively. This marker is applicable across hospitals, imaging protocols, and treatments, and has the potential to tailor neoadjuvant and adjuvant treatment at the individual level.Funding: This research was supported by the National Natural Science Foundation of China (grant numbers 82071885 and 81771802 and 81771893) and the National Youth Talent Support Program of China.Declaration of Interest: We declare no competing interests.Ethical Approval: IRB approval for the retrospective review of imaging and clinical data was obtained from the local ethics committees for all cohorts. The need for informed consent was waived.