Click here to

Session: Outcome Modeling and Assessment [Return to Session]

Predicting Osteoradionecrosis From Head and Neck Radiotherapy Using a ResidualĀ convolutional Neural Network

B Reber*, B Anderson, A Mohamed, L Van Dijk, B Rigaud, M McCulloch, Y He, M Woodland, C Fuller, S Lai, K Brock, UT MD Anderson Cancer Center, Houston, TX


TU-A-TRACK 6-4 (Tuesday, 7/27/2021) 10:30 AM - 11:30 AM [Eastern Time (GMT-4)]

Purpose: Osteoradionecrosis (ORN) is a detrimental radiation therapy (RT)-induced toxicity for patients after treatment of head and neck cancer. The ability to predict ORN before radiation treatment is performed could enable further optimization of the RT plan or stratification to advanced techniques (e.g. MR-LINAC, adaptive RT, dose de-escalation). A residual convolutional neural network is developed to predict ORN occurrence to aid in this endeavor.

Methods: A residual convolutional neural network was constructed using a hyperparameter search to find the model with optimum performance, including 2D and 3D architectures. Model input consisted of 3D computed tomography (CT) images, dose distributions, and mandible contours with size (32, 128, 128). Binary outcome of ORN is used for model output. The 148 subject data was split into 111 cases (40 ORN+) training and 37 cases (11 ORN+) validation. Positive cases were oversampled to ensure a match between ORN+ and ORN- in both datasets. Accuracy, sensitivity, specificity, and area under the receiver operator characteristic curve (AUC) are used for model evaluation. Results were compared with logistic regression models using a 0.8/0.2 train/test split and 10-fold cross-validation.

Results: The best logistic regression models had 0.68 AUC using only the dose distribution (V60) and 0.70 AUC using the dose distribution and clinical variables (N_stage+V65). The best deep learning model had 64 starting filters, 5 down-sampling operations and 1 residual block in-between down-sampling operations. The model was a 3D architecture using CT, dose distribution, and mandible contour and had 0.79 accuracy, 0.85 recall, 0.76 precision, 0.83 AUC. 3D deep learning architectures outperformed 2D architectures (0.53 AUC).

Conclusion: The model predicts ORN with higher performance metrics compared to logistic regression models. Further evaluation on multi-institutional test data is needed. The model could be implemented during the treatment planning stages to aid in patient decision making and treatment stratifications.



    Radiation Risk, Image Analysis, Dose Response


    IM/TH- Image Analysis Skills (broad expertise across imaging modalities): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)

    Contact Email