A comparative analysis of deep learning models for assisting in the diagnosis of periapical lesions in periapical radiographs.

PURPOSE

Numerous studies have investigated the use of convolutional neural network (CNN) models for detecting periapical lesions(PLs). However, limited research has focused on evaluating their potential in assisting clinicians with diagnosis. This study aims to utilize two deep learning(DL) models, ConvNeXt and ResNet34, to aid novice dentists in the detection of PLs on periapical radiographs (PRs). By assessing the diagnostic support provided by these models, this research seeks to promote the clinical application of DL in dentistry.

MATERIALS AND METHODS

In this study, 1,305 PRs were gathered and then split into a training set of 1,044 images and a validation set of 261 images, following an 80/20 ratio. The model's effectiveness was assessed using various measures, including precision, sensitivity, F1 score, specificity, accuracy, and the area under the curve (AUC). To evaluate the impact of the model on diagnostic performance by novice dentists, we used an additional set of 800 individual teeth PRs, which were not included in the training or validation sets. The diagnostic performance and time of three novice dentists were measured both with and without model assistance.

RESULTS

The precision of ConvNeXt was 85.93%, with an F1 score of 0.92, accuracy of 91.25%, sensitivity of 98.49%, specificity of 84.11%, and an AUC of 0.9693, outperforming ResNet34 across all metrics. In comparison, ResNet34 achieved a precision of 83.08%, an F1 score of 0.84, accuracy of 81.63%, sensitivity of 84.38%, specificity of 78.13%, and an AUC of 0.8988. In the model-assisted diagnosis phase, both ConvNeXt and ResNet34 improved the diagnostic performance of novice dentists. With the help of ConvNeXt, the average AUC of three dentists increased from 0.88 to 0.94, while with ResNet34, the average AUC of the three dentists improved from 0.88 to 0.91. ConvNeXt performed better than ResNet34 (p < 0.05). Additionally, ConvNeXt reduced the average diagnostic time of the three dentists from 178.8 min to 141.9 min, while ResNet34 reduced the average diagnostic time from 178.8 min to 153.6 min.

CONCLUSION

ConvNeXt significantly improved the diagnostic performance of novice dentists and reduced the time required for diagnosis, thereby enhancing clinical efficiency in both diagnosis and treatment. This model shows potential for application in dental clinics or educational institutions where experienced specialists are limited, but there is a large presence of novice, less-experienced dentists.