Deep Learning-Based Detection of Periapical Lesions in Periapical Radiographs

PurposePeriapical radiography is an effective method for detection of periapical lesions. However, owing to individual expertise, manually identifying and classifying lesions in images can be time-consuming, labor-intensive, and subjective. This study explored the feasibility of using the deep learning model ResNet34 to assist young dentists in diagnosing periapical lesions in periapical radiographs and evaluated the model's performance in identifying lesions of different severities.MethodsA total of 1,000 periapical radiographs were included in this study, encompassing a validation set of 200 images. The overall performance of the three young dentists in identifying periapical lesions, with and without the aid of the ResNet34 deep learning model, was compared using the evaluation criteria of accuracy, recall, precision, and F1 score. Additionally, three experienced dentists classified periapical lesion-positive cases into mild and severe cases to evaluate the model's performance in diagnosing mild and severe periapical lesions.ResultsThree young dentists exhibited an accuracy of 91.57%, a recall of 90.83%, a precision of 92.24%, and an F1 score of 0.91 in identifying periapical lesions without auxiliary diagnosis. When auxiliary diagnosis was employed, their accuracy increased to 95.01%, recall increased to 93.89%, precision increased to 96.09%, and F1 score increased to 0.94. The overall accuracies of the ResNet34 model for identifying mild and severe periapical lesions were 32.8% and 98.4%, respectively.ConclusionThis study demonstrated that deep learning-assisted diagnosis can significantly enhance the performance of young dentists in diagnosing periapical lesions on periapical radiographs. Notably, the ResNet34 model exhibited a superior performance in identifying severe periapical lesions.