Artificial intelligence in tongue diagnosis: classification of tongue lesions and normal tongue images using deep convolutional neural network.

OBJECTIVE: This study aims to classify tongue lesion types using tongue images utilizing Deep Convolutional Neural Networks (DCNNs). METHODS: A dataset consisting of five classes, four tongue lesion classes (coated, geographical, fissured tongue, and median rhomboid glossitis), and one healthy/normal tongue class, was constructed using tongue images of 623 patients who were admitted to our clinic. Classification performance was evaluated on VGG19, ResNet50, ResNet101, and GoogLeNet networks using fusion based majority voting (FBMV) approach for the first time in the literature. RESULTS: In the binary classification problem (normal vs. tongue lesion), the highest classification accuracy performance of 93,53% was achieved utilizing ResNet101, and this rate was increased to 95,15% with the application of the FBMV approach. In the five-class classification problem of tongue lesion types, the VGG19 network yielded the best accuracy rate of 83.93%, and the fusion approach improved this rate to 88.76%. CONCLUSION: The obtained test results showed that tongue lesions could be identified with a high accuracy by applying DCNNs. Further improvement of these results has the potential for the use of the proposed method in clinic applications.

Dental implant brand and angle identification using deep neural networks.

STATEMENT OF PROBLEM: Determining the brand and angle of an implant clinically or radiographically can be challenging. Whether artificial intelligence can assist is unclear. PURPOSE: The purpose of the present study was to determine the brand and angle of implants from panoramic radiographs with artificial intelligence. MATERIAL AND METHODS: Panoramic radiographs were used to classify the accuracy of different dental implant brands through deep convolutional neural networks (CNNs) with transfer-learning strategies. The implant classification performance of 5 deep CNN models was evaluated using a total of 11 904 images of 5 different implant types extracted from 2634 radiographs. In addition, the angle of implant images was estimated by calculating the angle of 2634 implant images by applying a regression model based on deep CNN. RESULTS: Among the 5 deep CNN models, the highest performance was obtained in the Visual Geometry Group (VGG)-19 model with a 98.3% accuracy rate. By applying a fusion approach based on majority voting, the accuracy rate was slightly improved to 98.9%. In addition, the root mean square error value of 2.91 degrees was obtained as a result of the regression model used in the implant angle estimation problem. CONCLUSIONS: Implant images from panoramic radiographs could be classified with a high accuracy, and their angles estimated with a low mean error.

An efficient 3D column-only P300 speller paradigm utilizing few numbers of electrodes and flashings for practical BCI implementation.

The event related P300 potentials, positive waveforms in electroencephalography (EEG) signals, are often utilized in brain computer interfaces (BCI). Many studies have been carried out to improve the performance of P300 speller systems either by developing signal processing algorithms and classifiers with different architectures or by designing new paradigms. In this study, a new paradigm is proposed for this purpose. The proposed paradigm combines two remarkable properties of being a 3D animation and utilizing column-only flashings as opposed to classical paradigms which are based on row-column flashings in 2D manner. The new paradigm is utilized in a traditional two-layer artificial neural networks model with a single output neuron, and numerous experiments are conducted to evaluate and compare the performance of the proposed paradigm with that of the classical approach. The experimental results, including statistical significance tests, are presented for single and multiple EEG electrode usage combinations in 1, 3 and 15 flashing repetitions to detect P300 waves as well as to recognize target characters. Using the proposed paradigm, the best average classification accuracy rates on the test data are improved from 89.97% to 93.90% (an improvement of 4.36%) for 1 flashing, from 97.11% to 98.10% (an improvement of 1.01%) for 3 flashings and from 99.70% to 99.81% (an improvement of 0.11%) for 15 flashings when all electrodes, included in the study, are utilized. On the other hand, the accuracy rates are improved by 9.69% for 1 flashing, 4.72% for 3 flashings and 1.73% for 15 flashings when the proposed paradigm is utilized with a single EEG electrode (P8). It is observed that the proposed speller paradigm is especially useful in BCI systems designed for few EEG electrodes usage, and hence, it is more suitable for practical implementations. Moreover, all participants, given a subjective test, declared that the proposed paradigm is more user-friendly than classical ones.