Dynamic Prediction of Toxicities in Head and Neck Cancer Radiotherapy by 3D Convolutional Neural Network Using Daily Cone-Beam CTs.

PURPOSE/OBJECTIVE(S): Radiotherapy (RT) is essential in head and neck cancer (HNC) treatments, but often causes significant toxicity. Different machine learning models have shown promise in predicting RT-induced toxicity, but none have yet integrated the fluctuating anatomical changes. By integrating daily cone-beam CTs (CBCT) allowing sequential anatomical views, our aim is to build a dynamic predictive model for three major HNC RT toxicities: reactive feeding tube placement, hospitalization and radionecrosis (RN).

MATERIALS/METHODS: 292 HNC cases treated with curative RT between 2017 and 2019 at our institution were retrospectively analyzed for clinical and radiological data. VoxelMorph, a deep deformable registration model, integrated the daily anatomical deformations between each CBCT and the planning CT, then converted them to Jacobian determinant matrix (Jf ). Resnet, a convolutional neural network with multiple layers was trained using a 5-fold cross validation to integrate both radiological and clinical data. Each toxicity was classified as a binary decision using the cross-entropy loss to account for a class imbalance. Its predictive performance was compared to the baseline model using only clinical data.

RESULTS: The cohort included 78% men and 22% women, with a median age of 63 years (range 35-84). Primary cancer sites were 46% oropharynx, 19% larynx, 14% oral cavity, 7.5% nasopharynx, 5% hypopharynx, 4% unknown primary and 5% others; and stage ranged between Tx-4b N0 and 3b M0 (AJCC 8th Ed). Induction chemotherapy, concurrent chemotherapy, and adjuvant RT was used in 9%, 57% and 20% of patients, respectively. The incidence of feeding tube, hospitalization and RN was 19.9%, 7.2%, and 3.8%, respectively. Integrating Jf from the 10th RT CBCT showed better accuracy for each toxicity prediction: feeding tube (69.1% > 57.2%), hospitalization (75.3% > 63.1%) and RN (85.8% > 75.7%). Integrating both the raw CBCT and Jf improved hospitalization prediction (79.0% > 73.6%). Substituting Jf for the raw CBCT improved the prediction for RN (79.7% > 74.7%) and hospitalization (73.6% > 64.4%). For feeding tube, predictive performance of the Jf model trained against deformations showed a positive correlation between its performance and the RT received (r2 > 0.9) with increasing RT fractions, with a maximum accuracy of 83.1% at the 25th fraction. No such correlation was found for RN or hospitalization prediction.

CONCLUSION: To our knowledge, this is the first study showing promising results to predict HNC RT toxicities using daily per-treatment CBCT. Next steps involve integrating both the radiomic and the dosimetric inputs to build a more powerful model. This could expand to predict therapeutic outcomes and, ultimately, could guide decisions in individualized RT.