Termografia infravermelha e inteligência artificial como auxiliares no diagnóstico da disfunção temporomandibular

Abstract

Temporomandibular disorders (TMD) are defined as set of musculoskeletal and neuromuscular conditions involving the temporomandibular joints (TMJ), masticatory muscles and associated structures. The RDC/TMD (Research diagnostic criteria for temporomandibular disorders) is the main and most accepted diagnostic tool for TMD. However, the application of this instrument is time consuming and demands professionals training. In cases of muscle disorders, infrared thermography (IT) has shown positive results by providing information regarding local microcirculation. The use of artificial intelligence techniques can help in the evaluation of thermograms and contribute to the diagnosis of TMD. This study assessed three methods of attribute extraction machine learning (ML) techniques: radiomic, semantic and the association of radiomic and semantic features on TMD detection using IT and palpation pain data; and to determine which ML classification algorithm, KNN, SVM e MLP, is the most efficient for this purpose. This is a cross-sectional observational study, approved by the research ethics committee protocol CAAE: 73417017200005187. The sample of this study was composed of 41 patients with TMD, selected using the RDC/TMD,and 37 pacients without TMD, volunteers who had the absence of TMD confirmed by the Fonseca anamnestic index. The selected patients had 18 to 60 years of age, of both sexes. All patients who had systemic changes that could interfere with thermograms assessment were excluded from the study. The thermographic exams were carried out in the infrared thermography laboratory of the State University of Paraíba following the protocols of the American Academy of Thermology, using the FLIR T650 infrared thermal camera, mounted on a tripod at 0.8m from the patient's face. Two thermograms in lateral norm (right and left side) of each patient were acquired. To analyze the temperature data, obtained using the FLIR Tools v.6.4 software, an observer using the circle tool obtained the temperatures in two regions in the anterior temporal muscle and in six regions in the masseter muscle. Temperature and pain on palpation data were tabulated and used as semantic features. To extract radiomic features, Fiji Image J software (version 1.8.0) was used to segment the regions of interest (ROI) corresponding to the masseter and anterior temporal muscles. After segmentation, twenty (radiomic) texture attributes, ten per assessed muscle, were extracted through co-occurrence matrices at the standard 0º angle: contrast, correlation, energy, homogeneity, entropy, trimmed mean, kurtosis, asymmetry, deviation pattern and variance. Subsequently, the texture attributes underwent dimensionality reduction procedure through principal component analysis. Using the ML classification algorithms: KNN, SVM and MLP, the accuracy, precision, sensitivity, and specificity were calculated for the associated radiomic, semantic and radiomic-semantic attribute extraction methods for the diagnosis of TMD. Data were statistically analyzed using the Hopkins test for cluster evaluation, ANOVA test for repeated measures and post Tukey test. The accuracy values of the training sample and the test sample only differed statistically for the association of radiomic-semantic attribute extraction methods (p<0,05). The MLP classifier algorithm showed the best accuracy values, differing from KNN and SVM MLP for the radiomic-semantic association method (p<0,05). The accuracy, precision and sensitivity values of the semantic attributes and the radiomic-semantic association differed statistically from the radiomic attributes (p<0,05). The semantic attributes had the best specificity values and differed statistically from the radiomic attributes (p<0,05). The semantic attribute extraction method and the association of radiomic and semantic attributes method showed better results than the radiomic method for the detection of TMD using IT and pain scale data. The MLP classifier algorithm must be used for this purpose when using radiomic and semantic attributes extraction methods. The use of AI associated with IT presents promising results for the detection of TMD.