Deep Learning for Automated Outcome Prediction in Oropharyngeal Cancer from Tumor and Lymph Node Imaging Data

Abstract

<h3>Purpose/Objective(s)</h3> Oropharyngeal carcinoma (OPC) is a malignancy with highly variable prognosis. Improved biomarkers and risk-stratification tools are needed to guide personalized treatment paradigms. Deep learning (DL) has shown significant promise in its ability to synthesize computed tomography (CT) imaging data for risk-stratification. However, thus far, CT models for OPC prognosis have not leveraged combined primary tumor and lymph node data for outcome prediction. In this study, we hypothesized that a CT-based, combined tumor and lymph node DL model could be developed and validated to predict survival in patients with OPC treated with radiotherapy (RT) based treatment. <h3>Materials/Methods</h3> Tumor and lymph node gross tumor volume contours from pretreatment CT scans for 1,033 OPC patients treated with definitive RT from two institutions from 2003-2013 were curated with linked survival data via The Cancer Imaging Archive. The combined primary and nodal contours with underlying CT data served as input for model training (70%, 743), tuning (10%, 83), and blinded test sets (20%, 207). We trained a convolutional neural network, adopted from ResNet101, customized with cause-specific Logistic Hazard loss functions to predict individual overall survival (OS). Multivariable Cox regression was conducted using clinical variables alone (AJCC 7^th Edition Stage, HPV-status, smoking [never, former, current], sex, and age) and combined with DL survival probability (DL-score). Primary endpoint was Concordance Index (C-index) for OS on the test set. Kaplan–Meier curves with log-rank test comparisons were generated for OS. <h3>Results</h3> Among 1,033 patients (median age: 59 years), 1.4%, 4.8%, 13.7% and 80.1% were Stage I, II, III and IV, respectively, 52.7%, 12.9% and 34.4% were HPV+, HPV-, and unknown, respectively, and 66.3% and 30.1% were treated with chemoradiotherapy and radiotherapy alone, respectively. Median follow-up time was 72.4 months. DL-score alone yielded comparable OS prediction compared to the clinical Cox model overall (C-index: 0.70 vs 0.68), and between HPV+ (0.66 vs 0.61) and HPV- (0.77 vs 0.76) patient subgroups. Combining DL-score and clinical variables into the Cox model yielded the highest predictive performance for OS (C-index: 0.74). DL-score alone stratified test set patients overall into three risk groups (low: n=77, intermediate: n=100, and high: n=30) with significant survival curve separation (<i>P</i><0.001), and 5-year OS of 90.9%, 71.0% and 50.0%, respectively. For the HPV+ subgroup (n=114), DL-score also yielded low, intermediate, and high-risk grouping with 5-year OS of 90.7%, 71.9%, and 64.3% (<i>P</i>=.32). <h3>Conclusion</h3> We developed and validated an imaging-based DL survival prediction model for patients with OPC treated with RT that prognosticates well, irrespective of HPV status. Performance is improved further by incorporating HPV and clinicopathologic factors. By improving prognostication, DL could be used to inform trial eligibility and personalized decision-making.