Automatic identification of dental implant brands with deep learning algorithms

Objevtive: The aim was to classify various dental implant brands using deep learning algorithms on panoramic radiographs as a way to reduce problems arising from the inability to determine the brand of dental implants. Method: Images from a total of 5375 cropped panoramic radiographs belonging to four different dental implant systems (NucleOSS, Medentika, Nobel, and Implance) were used. To enhance the images and reduce blurriness, the Contrast Limited Adaptive Histogram Equalization (CLAHE) filter was applied. The deep learning algorithms GoogleNet, ResNet-18, VGG16, and ShuffleNet were employed to classify the four different dental implant systems. To evaluate the classification performance of the algorithms, ROC curves and confusion matrices were generated. Based on these confusion matrices, accuracy, precision, recall, and F1 scores were calculated. The Z-test was used to compare the results of the performance metrics across the algorithms. Results: The accuracy rates of the deep learning algorithms were obtained as 96.00% for GoogleNet, 84.40% for ResNet-18, 98.90% for VGG16, and 84.80% for ShuffleNet. There was a statistically significant difference in accuracy between the VGG16 algorithm and the GoogleNet, ShuffleNet, and ResNet-18 algorithms (p<0.001, p<0.001, and p<0.001, respectively). Upon examining the ROC curves, it was found that the ROC curves for all four algorithms were close to the (0,1) point, indicating good model performance. Analysis of the confusion matrices revealed that the NucleOSS and Medentika brands were the most frequently confused dental implant brands across all algorithms. Conclusion: The high accuracy rates suggest that deep learning algorithms are considered a valuable and powerful method for potential use in determine the brands of dental implant. Keywords: dental implant, deep learning, automated identification