Dose Aware Toxicity Prediction in Head and Neck Cancer Patients Using a Deformable 3D CNN on Daily CBCT Acquisitions

With recent advancements in intensity-modulated radiotherapy and image guidance for cancer treatment, there is growing interest in Deep Learning-based Adaptive Radiotherapy due to its potential to mitigate radiation toxicity. During the course of treatment, daily acquisitions of cone beam computed tomography (CBCT) allows to monitor significant anatomical changes around the tumour target area and the response to treatment. Previous works have demonstrated the benefits of using anatomical deformation features as predictors of early toxicities. The objective of this work is to investigate the use of radiomic distributions to predict early reactive na-sogastric tubing (NG tube), radionecrosis and broad hospitalisation based on initial dosimetry plans. We propose a method combining inter-fractional anatomical deformation, dosimetry and clinical information to improve the prediction performance. For this work we implement a deformable registration pipeline and train a 3D convo-lutional neural network model on the Jacobian determinants of the deformation vector fields between different fractions of treatment. We exploit the dose plans as feature maps to focus the attention of the network on areas susceptible to radiation toxicity. We obtain balanced accuracy scores of 75.4% for radionecrosis at fraction 20,61.1% for hospitalisation after the first week of treatment and 74.1% for NG tube insertion after the 5
^th
week.