AI-Based Detection of Periapical Lesions in Endodontic Radiographs: a Diagnostic Accuracy Study

Periapical lesions pose a serious diagnosis problem in endodontics and there is high inter-observer variability in the conventional interpretation of radiographs. It has a solution in artificial intelligence (AI) technologies, which will improve the accuracy and consistency of diagnosis. This paper set out to determine the diagnostic accuracy of an AI system based on deep learning in identifying periapical lesions in endodontic radiographs as compared to clinical assessment and histopathological confirmation by an expert.

A retrospective diagnostic accuracy study on 1,247 periapical radiographs of patients who had endodontic treatment or periapical surgery with histopathological verification was done. The architecture of a convolutional neural network (CNN) with EfficientNetB4 was trained using 80% of the dataset, where 10% was to be used as a validation and 10% as a test set. The diagnostic performance was assessed against the histopathological findings as the reference standard. The same radiographs were independently evaluated by three experienced endodontists. Sensitivity and specificity, accuracy, positive predictive value (PPV) and negative predictive value (NPV) as well as area under the receiver operating characteristic curve (AUC-ROC) were determined.

The AI model had a total accuracy of 91.2 [standard deviation (SD) 1.7], a sensitivity of 93.4 (SD 1.5) and a specificity of 88.6 (SD 2.1). Endodontists had an average of 78.6 ± 4.3, which was highly lesser than the AI model (p <0.001). Agreement among clinicians (between different observers) was moderate (0.612). The AI model was superior in all categories of lesion sizes, showing significant improvement in cases of small lesions (less than 3 mm), with an accuracy of 87.8 compared to that of clinicians (64.2) (p<0.001).

The model of AI built has shown better accuracy in detection of periapical lesions when compared with expert clinical assessment, especially when small lesions were involved. Such AI-based diagnostic tools may be important in facilitating endodontic diagnosis and treatment planning.