Lung Cancer Recurrence Risk Prediction through Integrated Deep Learning Evaluation

SIMPLE SUMMARY: Few significant advances have been made over recent decades in predicting lung cancer progression risk after complete surgical removal of tumor in stage IA non-small-cell lung cancers (NSCLCs). Although several biomarkers have shown some predictive value, it is unclear whether these markers add value to traditional TNM staging. We developed an integrated deep learning evaluation (IDLE) score to combine patient’s preoperative lung CT image findings and postoperative pathologic assessment and found that this score can better predict cancer progression risk than TNM staging and tumor grade. Improved predictive value of the IDLE score was primarily due to the complementary use of tumor measurements in CT images from an entire lung as well as microscopic tissue characteristics. Our findings suggest that integrating measurements from different aspects of tumor morphology is more robust for increasing prediction accuracy than building on the measurements of similar aspects of tumor morphology. ABSTRACT: Background: Prognostic risk factors for completely resected stage IA non-small-cell lung cancers (NSCLCs) have advanced minimally over recent decades. Although several biomarkers have been found to be associated with cancer recurrence, their added value to TNM staging and tumor grade are unclear. Methods: Features of preoperative low-dose CT image and histologic findings of hematoxylin- and eosin-stained tissue sections of resected lung tumor specimens were extracted from 182 stage IA NSCLC patients in the National Lung Screening Trial. These features were combined to predict the risk of tumor recurrence or progression through integrated deep learning evaluation (IDLE). Added values of IDLE to TNM staging and tumor grade in progression risk prediction and risk stratification were evaluated. Results: The 5-year AUC of IDLE was 0.817 ± 0.037 as compared to the AUC = 0.561 ± 0.042 and 0.573 ± 0.044 from the TNM stage and tumor grade, respectively. The IDLE score was significantly associated with cancer recurrence (p < 0.0001) even after adjusting for TNM staging and tumor grade. Synergy between chest CT image markers and histological markers was the driving force of the deep learning algorithm to produce a stronger prognostic predictor. Conclusions: Integrating markers from preoperative CT images and pathologist’s readings of resected lung specimens through deep learning can improve risk stratification of stage 1A NSCLC patients over TNM staging and tumor grade alone. Our study suggests that combining markers from nonoverlapping platforms can increase the cancer risk prediction accuracy.