Artificial intelligence for detection of periapical lesions on intraoral radiographs: Comparison between convolutional neural networks and human observers

Objective The aim of this study was to compare the diagnostic performance convolutional neural networks (CNNs) with human observers for detection simulated periapical lesions on radiographs. Study Design Ten sockets were prepared in bovine ribs. Periapical defects 3 sizes sequentially created. radiographs acquired each socket no lesion and size a photostimulable storage phosphor system. Radiographs evaluated filter 6 image settings. A CNN architecture set up using Keras-TensorFlow. Separate CNNs randomly sampled training/validation data split by (5-fold cross-validation) (7-fold cross-validation). validation compared that oral radiologists sensitivity, specificity, area under receiver operating characteristic curve (ROC-AUC). Results Using random sampling, showed perfect accuracy data. When socket, mean ROC-AUC values 0.79, 0.88, 0.86, respectively; when filter, they 0.87, 0.98, 0.93, respectively. For radiologists, 0.58, 0.83, 0.75, Conclusions show promise detection. pretrained model yielded can be used further training larger samples and/or clinical