Development of an automatic detection model using artificial intelligence for the detection of cyst-like radiolucent lesions of the jaws on panoramic radiographs with small training datasets

Abstract

Recent reports have indicated that further standardization of medical care can be expected by implementing a double-check system using a deep convolutional neural network (DCNN). Panoramic radiograph is the standardized imaging modality that is routinely performed in dental practice and allows examination of the oral and maxillofacial region on a single image. We report the development of an artificial intelligence (AI) system for detecting cyst-like radiolucent lesions of the jaws on panoramic radiographs with small training datasets. We used 7160 panoramic radiographs showing cyst-like radiolucent lesions of the jaws as the data to train a deep learning algorithm using transfer learning and to construct a DCNN that would automatically detect the lesions. Next, we used 100 panoramic radiographs showing cyst-like radiolucent lesions as testing data and set four evaluation target areas: (1) an area from the maxillary anterior teeth to the premolars, (2) an area of the mandibular body (including chin), (3) an area of the right mandibular ramus, and (4) an area of the left mandibular ramus. The DCNN was evaluated on 400 target sites. By using deep learning with panoramic radiographs, we constructed an AI system based on a DCNN that could detect cyst-like radiolucent lesions of the jaws. The DCNN showed an accuracy of 98.3%, sensitivity of 94.4%, specificity of 99.7%, precision of 99.0%, recall of 94.4%, and F-score of 0.966. Our results suggest that this AI system may contribute to diagnostic support in future clinical practice.