Radiomics-based classification of medication-related osteonecrosis of the jaw using panoramic radiographs

Comprehensive Summary

This study attempts to estimate the likelihood of medication-related osteonecrosis of the jaw (MRONJ) development using an AI model with radiograph data. To train the model, radiographic images from 32 MRONJ patients and 57 non-MRONJ patients at Hiroshima University Hospital were used. After segmenting the radiographs, the researchers ended up with 13 shapes, 18 histograms, 75 textures and 744 wavelet features. The least absolute shrinkage and selection operator (LASSO) was used to shrink the data down to 10 key feature. Two machine learning models were trained using support vector machine (SVM) and neural network of multilayer perceptron (MLP). SVM achieved a sensitivity of 0.667, specificity of 0.833 and AUC (Area under the Curve) of 0.903. Similarly, MLP achieved a sensitivity of 0.833, specificity of 0.750 and AUC (Area under the Curve) of 0.903. Therefore, the models have the capability of distinguishing future MRONJ and non-MRONJ patients.

Outcomes and Implications

Medication-related osteonecrosis of the jaw (MRONJ) is generally caused by bone resorption inhibitors. Despite the low incidence of only 0.06% of patients, the intervention required in the advanced stages of the disease often require surgery, impacting one’s post treatment quality of life. Early MRONJ detection can potentially help patients to avoid serious intervention or degradation in their quality of life. This can be achieved efficiently with the use of automatic detection systems using artificial intelligence which would require less time for physicians to analyze the data. The physicians can therefore tailor medications for high risk MRONJ patients when performing surgery.