Identification and classification of dental implant systems using various deep learning-based convolutional neural network architectures

Background

Deep learning methods, particularly deep convolutional neural networks (CNNs), are likely to be one of the most transformative technologies for medical applications.

Aim/Hypothesis

The aim of the current study was to evaluate the efficacy of three major deep CNN algorithms for the identification and classification of dental implant systems.

Material and Methods

A total of 3,000 cropped panoramic and periapical radiographic images from three brands of implant systems óTSIII® SA, Superline®, and SLActive® bone level taperedó were divided into a training and validation dataset (n = 2400 [80%]) and a test dataset (n = 600 [20%]). Three pre-trained deep CNN architectures óVGG-19, GoogLeNet Inception-v3, and ResNet-50ó were used for preprocessing and transfer learning. The test datasets were assessed for accuracy, sensitivity, specificity, receiver operating characteristic (ROC) curve, area under the ROC curve, and confusion matrix.

Results

We demonstrated that the three major deep CNN architectures provided reliable performance (VGG-19- AUC = 0.891, 95% confidence interval [CI] 0.839-0.930 + Inception-v3- AUC = 0.922, 95% CI 0.876–0.955 + and ResNet-50- AUC = 0.907, 95% CI 0.858–0.944) with no statistically significant difference in terms of performance between the three architectures in identification and classification using radiographic images of three brands with similar internal-type implant systems.

Conclusion and Clinical Implications

This study demonstrated that the three major deep CNN architectures were useful for the identification and classification of implant systems using a small number of radiographic images.