Multi-Institutional Deep Learning Modeling of Mandibular Osteoradionecrosis Using Clinical and Dosimetric Data

https://doi.org/10.5281/zenodo.17487161

Abstract

Background: Mandibular osteoradionecrosis (ORN) is a severe late toxicity of head and neck radiotherapy (RT) characterized by non-healing necrotic bone, with an estimated incidence of 4-8% in the modern IMRT era [1]. Multiple patient, tumor, and treatment factors (e.g., dental extractions, tumor proximity, smoking, and high radiation dose) all contribute to ORN [1,2]. Predictive models that quantify individualized ORN risk could help guide preventive strategies, but there were no validated normal tissue complication probability (NTCP) models for ORN until recently [3]. This study creates a multi-institutional deep learning (DL) model that uses 3D dose distributions ("dosiomic" data) and clinical variables to predict mandibular ORN, with the hypothesis that a large, diverse dataset will improve predictive performance and generalizability.

Methods: A retrospective multi-institutional cohort of 1,184 head and neck cancer patients (389 ORN cases) from seven centers was created [4]. Clinical, demographic, and dosimetric information (3D mandibular dose-volume maps and dose-volume histogram metrics) was gathered. The mandibular dose distribution was processed using late fusion convolutional neural network (DenseNet architecture), with clinical risk factors integrated via late fusion. Model training applied 5-fold cross-validation to 80% of the pooled data and was tested on an internal test set (20%) and an external hold-out institution. Performance was evaluated using the area under the ROC curve (AUC), calibration, and F1-score.

Results: ORN patients received significantly higher mandibular radiation doses (median D30% and V70Gy), as well as a higher rate of pre-RT dental extractions and smoking [4]. The multi-modal DL model had an internal test AUC of ~0.72 and good calibration, similar to a logistic regression model (AUC ~0.69) [3]. In an independent external cohort, performance was lower (AUC ~0.60), indicating institutional heterogeneity [3]. In terms of discrimination and calibration, a late-fusion DL model performed slightly better than early-fusion and single-modality models [3].

Conclusions: This large-scale study demonstrates the feasibility of predicting ORN risk with multi-institutional data and deep learning. While the DL model performed similarly to traditional models, it offers spatial dose-response insights. Key predictors included mandibular dose-volume metrics (D30%, V70Gy) and dental risk factors [4]. Further research will concentrate on incorporating advanced architectures (e.g., transformers) and addressing domain shifts in order to improve generalizability. This approach shows promise for personalized ORN risk stratification, allowing for tailored prophylactic dental care or adaptive RT planning to reduce ORN in high-risk patients.