Comparing 3D deformations between longitudinal daily CBCT acquisitions using CNN for head and neck radiotherapy toxicity prediction

• DOI: 10.48550/arxiv.2303.03965
• Publication Date: 2023-01-01
• Authors: William Le; Chulmin Bang; Philippine Cordelle; Daniel Markel; Phuc Félix Nguyen...; Houda Bahig; Samuel Kadoury
• Keywords: FOS: Computer and information sciences; Computer Science - Machine Learning; 03 medical and health sciences; 0302 clinical medicine; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition; 3. Good health; Machine Learning (cs.LG)

Adaptive radiotherapy is a growing field of study in cancer treatment due to it's objective sparing healthy tissue. The standard care several institutions includes longitudinal cone-beam computed tomography (CBCT) acquisitions monitor changes, but have yet be used improve tumor control while managing side-effects. aim this demonstrate the clinical value pre-treatment CBCT acquired daily during radiation therapy for head and neck cancers downstream task predicting severe toxicity occurrence: reactive feeding tube (NG), hospitalization radionecrosis. For this, we propose deformable 3D classification pipeline that component analyzing Jacobian matrix deformation between planning CT CBCT, as well data. model based on multi-branch residual convolutional neural network, registration pair VoxelMorph architectures. Accuracies 85.8% 75.3% was found radionecrosis hospitalization, respectively, with similar performance early after first week treatment. NG risk, improves increasing timing fraction, reaching 83.1% $5_{th}$