Inteligência artificial como método auxiliar na estimativa de idade em radiografia panorâmica

Abstract

The knowledge of the age of the living corpses is of criminal importance to the illegal responsibility, as the disaster and the creation of the biological of so not identified in mass profile. Relatively non-invasive approach to the methodology used, well known as the Kvaal age method. However, estimating age through radiographic methods can present a high margin, in addition to error approaches, time and training of examiners. Thus, the possibility of increasing the accuracy of the estimate with the use of artificial intelligence that has been achieving results in several areas of knowledge. This study aimed to estimate age using the Kvaal radiographic method on radiographs using an auxiliary artificial intelligence. This was a cross-sectional observational study approved by the Research Ethics Committee under CAAE: 56455322.6.0000.5187. A total of 554 radiographs were selected from a private clinic from a database in northeastern Brazil collected between 2017 and 2019 and taken on Carestream® CS 8100 3D and Carestream® CS 900 3D equipment. The radiographs were of patients aged between 13 and 57 years with the left maxillary central incisor and left maxillary canine erupted and healthy. The intraobserver agreement of the Kvaal method was defined through a Kappa test, without which the flow of confirmation of an adequate pattern. To estimate age using the Kvaal method, linear measurements were performed on the images using ImageJ® software. In order to compensate for any difference in the magnification of the radiographic images, the proportions between the measurements performed on the tooth and the pulp cavity were calculated. Data from the Kvaal method were statistically analyzed using the Mann-Whitney test, the Spearman correlation coefficient, the Student's t test and linear regression with its respective coefficient of determination to estimate the ages and analyze the variability of data from the I estimated. The margin of error used in the decision of the statistical tests was of 5% and the intervals with 95% of confidence. Regarding age prediction through artificial intelligence, three approaches were used: semantics, radiomics and the concatenation of semantic and radiomic attributes. The sample partitioning was performed in the proportion 85/15 (85% of the sample for training and 15% for tests) in the three approaches. From the training, the regression algorithms with the best performances were selected and used in the test group: Logistic Regression, Linear Regression, Random Forest Reg, Gradient Boost Reg and XG Boosting Reg. The tabulated data of the Kvaal method were extracted, generating nineteen semantic attributes. The maxillary central incisors and maxillary canines were segmented from the panoramic radiographs manually using the GIMP® software and generated seven radiomic attributes (contrast, correlation, energy, homogeneity, entropy, mean and deviation) for each tooth, totaling fourteen attributes. Finally, the concatenation of the data resulted in thirty-three attributes (semantic and radiomic). To analyze the statistical results of machine learning, the Mean Absolute Error (EMA) was applied. The isolated Kvaal method is applicable in the studied population (EMA 5.68) with emphasis on the maxillary incisor (EPE 7.10) which obtained greater estimation precision than the canine (EPE 7.38). Regarding applied machine learning, the extraction of semantic and radiomic attributes presented similar performance in the complete base (EMA 5.23), while the concatenation of these attributes obtained superior precision in the estimation of ages (EMA 4.77). The XG Boosting Reg classifier showed promising performance. Machine learning showed higher accuracy (EMA 4.77) when estimating the ages of the population studied when compared to the Kvaal method (EMA 5.68). Therefore, artificial intelligence can be applied as an auxiliary approach in age estimation using panoramic radiographs.