Deep Learning Based Recurrence Prediction in Head and Neck Cancers after Radiotherapy

Purpose/Objective(s) Nearly half of patients with head and neck (H&N) cancers experience recurrence, yet the challenge of prognosticating high-risk individuals for intensified radiotherapy persists. Here we aimed to develop a deep learning (DL) model for predicting H&N cancer recurrence based on clinical and radiomic features (texture, shape, intensity) from radiation planning CT simulation scans. Materials/Methods We analyzed contrast enhanced CT scans of 249 patients with H&N cancers from The Cancer Imaging Archive (TCIA). This dataset included patients with known recurrence status, defined as a locoregional recurrence or distant metastasis. Radiation treatment clinical target volume (CTV) contours from the CT scans were extracted, z-score normalized, and resized to 128 × 128 × 128 volumes before analyzing radiomic features. For the recurrence prediction task, we utilized a modified 3D variant of Resnet50, with patients randomly assigned to training (n=199) or test (n=50) sets. ROC, AUC, sensitivity, and specificity were reported across 5-fold cross validation. Results The majority of patient in the study involved oropharyngeal cancer (68%) with a median age of 63 years and follow up of 42.9 months. In total, 26.5% of patients experienced a recurrence in a median time of 14.9 months. The combined clinical-radiomic DL model achieved an AUC of 0.75 with an accuracy of 0.71. When considering only the highest-performing clinical features (TNM stage, treatment modality, HPV status, and sex), the model yielded an AUC of 0.6 and an accuracy of 0.61. Finally, utilizing only the top 5 out of 1218 radiomic features in the model resulted in an AUC of 0.64 with an accuracy of 0.47. Conclusion This study underscores the potential of DL models in predicting H&N cancer recurrence by combining clinical and radiomic features from CT scans. The promising results of our integrated model suggest its role in identifying high-risk individuals who could benefit from intensified treatments, ultimately improving the management and outcomes for patients with H&N cancers. Even when using clinical and radiomic features separately, valuable insights into patient risk stratification were obtained. Further refinement and validation of these models have the potential to pave the way for more precise and personalized treatment strategies.