Development of a Multi-Modality Clinical Decision-Support System for Risk Stratification in Non-Small Cell Lung Cancer
Abstract
Surgery is the primary treatment option for patients with early-stage non-small cell lung cancer (NSCLC). However, it is difficult for clinicians to determine which patients are at a high risk of recurrence after surgery and who would benefit from adjuvant therapies. Computed tomography (CT) and positron emission tomography (PET) images are used to determine the cancer staging, which is the gold standard for predicting recurrence. Previous studies have shown that these images may hold prognostic information and can augment clinical data. Machine learning models (MLMs) built utilizing quantitative imaging information, which can outperform current clinical prognostic markers, would therefore aid in decision-making for NSCLC patients. The objective of this work was to develop, validate, and evaluate a clinical decision-support system (CDSS) integrating multi-modality MLMs utilizing imaging and clinical information for the risk stratification of NSCLC patients after surgery. A semi-automatic segmentation algorithm was developed to delineate the tumour on CT. The algorithm provided reproducible segmentations regardless of experience level and was packaged in an efficient and user-friendly interface. CT and PET radiomic features were then extracted from the tumour and peri-tumoural regions, along with vertebral bodies L3-L5 to assess bone marrow uptake on PET. AMLM using these radiomic features and clinical features achieved a concordance of 0.76 and outperformed stage when stratifying patients into high and low risks of recurrence. Subsequently, a deep learning model (DLM) was developed, trained, and externally validated using CT, PET, clinical, surgical, and pathological information from 500 NSCLC patients across two institutions. The DLM outperformed stage achieving an area-under-the-receiver-operating-curve (AUC) of 0.78 and 0.66 in the testing and external validation cohorts, respectively. The validated DLM was integrated into a CDSS, and a user-study was conducted to compare the performance of the DLM to clinicians and evaluate how the CDSS would be best adopted into clinical practice. The CDSS was able to augment the performance of clinicians when predicting post-surgical recurrence and improved their decision confidence. In conclusion, these contributions have the potential to lead to more personalized treatment decisions and improve outcomes for patients with NSCLC.