Improving the accuracy of publicly available search engines in recognizing and classifying dental visual assets using convolutional neural networks

Aim: To assess the accuracy of DigiBrain4, Inc (DB4) Dental Classifier and DB4 Smart Search Engine* in recognizing, categorizing, and classifying dental visual assets as compared with Google Search Engine, one of the largest publicly available search engines and the largest data repository. Materials and methods: Dental visual assets were collected and labeled according to type, category, class, and modifiers. These dental visual assets contained radiographs and clinical images of patients' teeth and occlusion from different angles of view. A modified SqueezeNet architecture was implemented using the TensorFlow r1.10 framework. The model was trained using two NVIDIA Volta graphics processing units (GPUs). A program was built to search Google Images, using Chrome driver (Google web driver) and submit the returned images to the DB4 Dental Classifier and DB4 Smart Search Engine. The categorical accuracy of the DB4 Dental Classifier and DB4 Smart Search Engine in recognizing, categorizing, and classifying dental visual assets was then compared with that of Google Search Engine. Results: The categorical accuracy achieved using the DB4 Smart Search Engine for searching dental visual assets was 0.93, whereas that achieved using Google Search Engine was 0.32. Conclusion: The current DB4 Dental Classifier and DB4 Smart Search Engine application and add-on have proved to be accurate in recognizing, categorizing, and classifying dental visual assets. The search engine was able to label images and reject non-relevant results. Keywords: dental visual assets, artificial intelligence, dental radiographs, dental clinical images, dental classifier, smart search engine, machine learning, deep learning, convolutional neural network