Development and Validation of a Deep Learning System with Tumor- and Patient-Centric Imaging Analysis to Improve Risk-Stratification in Oropharyngeal Cancer

• Published in: International journal of radiation oncology, biology, physics Vol. 120; no. 2; pp. e804 - e805
• Main Authors: Ye, Z., Ravipati, Y., Hoebers, F., Zapaishchykova, A., Zha, Y., Tak, D., Mahootiha, M., Pardo, J.C., Guthier, C.V., Tishler, R.B., Margalit, D.N., Haddad, R.I., Schoenfeld, J.D., Mak, R.H., Aerts, H., Kann, B.H.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01-10-2024
• ISSN: 0360-3016
• DOI: 10.1016/j.ijrobp.2024.07.1767

Oropharyngeal cancer (OPC) outcomes are heterogeneous, though management remains driven by stage of disease. Improved biomarkers may guide treatment approach. Deep learning (DL)-based tumor imaging analysis has been shown to improve risk stratification in OPC. Since then, patient-centric, imaging-derived biomarkers such as skeletal muscle area and adipose density have emerged. We hypothesized a CT-based, DL system that analyzes tumor- and patient-specific features would improve patient risk-stratification for OPC.
2,330 OPC patients treated with definitive radiotherapy from 1991 to 2018 (7 institutions) were used for model development. External validation sets were curated from 2 cancer centers (C1, n = 176; C2, n = 244). A multi-step pipeline was developed using U-Net-based segmentation of gross tumor and lymph nodes and C3 vertebrae-level skeletal muscle and adipose to calculate tumor volume, skeletal muscle area (SMA), and adipose density, respectively. A DL survival model was developed to predict DL-scores from raw imaging data. Cox proportional hazards models for overall survival (OS), recurrence-free survival (RFS), and distant metastasis-free survival (DMFS) were fit to investigate the incremental benefit of tumor- and patient-centric factors, starting with a clinical model benchmark (AJCC 7th edition T, N, and overall stage, HPV-status, smoking, sex, and age) with primary endpoint of Concordance Index (CI). K-Means clustering stratified patients into risk groups based on Cox model-derived probabilities.
Among 2,750 patients (median age: 60 years), 56% were HPV+, 14% HPV-, and 30% HPV-unknown. The integrated model with clinical variables, tumor and node volumetrics, DL-score, SMA, and adipose density, yielded the best performance and was stable across both external sets for OS, RFS, and DMFS (Table 1). The combined model improved risk-group stratification in both test sets with significant survival curve separation (p<0.001) for OS, PFS, and DMFS. For example, in C2, the combined model showed improved DMFS stratification with 100% (low-risk, n = 167), 89.1% (int-risk, n = 64), and 8.3% (high-risk, n = 12), compared to 96.9% (low-risk, n = 161), 86.1% (int-risk, n = 72), and 70% (high-risk, n = 10) using the clinical variable model.
We developed a CT-based DL system that predicts OPC risk via patient and tumor imaging analysis. The incorporation of DL imaging-based risk-stratification into clinical care may better inform treatment decision-making and clinical trial selection.