Australian dental journal · 2026
Machine Learning-Based Screening for Osteonecrosis of the Jaw in Panoramic Radiographs.
СтатьяОбзорНазвание и аннотация: Перевод пока не подготовлен
Обработка статьи · этапы, ошибки и управление
Загрузка состояния…
Ошибки и ограничения
История · было → стало
Состояние обновляется каждые 5 секунд, пока блок открыт. Процент показывается только там, где известен объём работы. Пауза останавливает продолжение этапа; текущий запрос к источнику может завершиться.
Nagyová Valéria, Blaňár Dominik, Kybic Jan, Tichý Antonín
Аннотация
BACKGROUND: Osteonecrosis of the jaw (ONJ) is a serious condition associated mainly with antiresorptive therapy and maxillofacial radiotherapy. As detection of its radiographic findings may facilitate patient referral, this study aimed to develop convolutional neural networks (CNNs) for ONJ screening in panoramic radiographs (OPGs).
METHODS: A dataset of 888 ONJ-positive OPGs and 888 ONJ-negative controls was assembled and annotated using bounding boxes. The dataset was split at patient level 70-10-20 for training, validation, and testing. Multiple CNNs (ResNet18, ResNet34, MobileNetV2, EfficientNet-B0) were trained on cropped image regions and evaluated using standard performance metrics. Performance on full OPGs (without cropping) was tested, and subgroup analysis was performed. Regions contributing to model predictions were visualised using Gradient-weighted Class Activation Mapping (Grad-CAM).
RESULTS: EfficientNet-B0 performed best, with per-crop accuracy, sensitivity, specificity, precision, and F1-score of 92.24%, 89.80%, 94.69%, 94.42% and 82.05%, respectively. At 0.9 probability threshold, per-image accuracy, sensitivity, specificity, precision, and F1-score equalled 76.65%. Grad-CAM highlighted clinically relevant regions; however, differentiation from other changes in bone density was suboptimal.
CONCLUSIONS: Despite promising results at crop level, model performance in full OPG analysis is currently insufficient for clinical application. Further optimisation of localisation strategies and external validation are needed for reliable clinical application.
METHODS: A dataset of 888 ONJ-positive OPGs and 888 ONJ-negative controls was assembled and annotated using bounding boxes. The dataset was split at patient level 70-10-20 for training, validation, and testing. Multiple CNNs (ResNet18, ResNet34, MobileNetV2, EfficientNet-B0) were trained on cropped image regions and evaluated using standard performance metrics. Performance on full OPGs (without cropping) was tested, and subgroup analysis was performed. Regions contributing to model predictions were visualised using Gradient-weighted Class Activation Mapping (Grad-CAM).
RESULTS: EfficientNet-B0 performed best, with per-crop accuracy, sensitivity, specificity, precision, and F1-score of 92.24%, 89.80%, 94.69%, 94.42% and 82.05%, respectively. At 0.9 probability threshold, per-image accuracy, sensitivity, specificity, precision, and F1-score equalled 76.65%. Grad-CAM highlighted clinically relevant regions; however, differentiation from other changes in bone density was suboptimal.
CONCLUSIONS: Despite promising results at crop level, model performance in full OPG analysis is currently insufficient for clinical application. Further optimisation of localisation strategies and external validation are needed for reliable clinical application.