Convolutional neural network for automated classification of osteonecrosis and related mandibular trabecular patterns

The present study aimed to develop and validate a tool for the automated classification of normal, affected, and osteonecrosis mandibular trabecular bone patterns in panoramic radiographs using convolutional neural networks (CNNs).

A dataset of 402 panoramic images from 376 patients was selected, comprising 112 control radiographs from healthy patients and 290 images from patients treated with antiresorptive drugs (ARD). The latter was subdivided in 70 radiographs showing thickening of the lamina dura, 128 with abnormal bone patterns, and 92 images of clinically diagnosed osteonecrosis of the jaw (ONJ). Four pre-trained CNNs were fined-tuned and customized to detect and classify the different bone patterns. The best performing network was selected to develop the classification tool. The output was arranged as a colour-coded risk index showing the category and their odds. Classification performance of the networks was assessed through evaluation metrics, receiver operating characteristic curves (ROC), and a confusion matrix. Furthermore, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visualise class-discriminative regions.

All networks correctly detected and classified the mandibular bone patterns with optimal performance metrics. InceptionResNetV2 showed the best results with an accuracy of 96 %, precision, recall and F1-score of 93 %, and a specificity of 98 %. Overall, most misclassifications occurred between normal and abnormal trabecular bone patterns.

CNNs offer reliable potentials for automatic classification of abnormalities in the mandibular trabecular bone pattern in panoramic radiographs of antiresorptive treated patients.

A novel method that supports clinical decision making by identifying sites at high risk for ONJ.