Abordagem híbrida com rede neural convolucional e modelos de regressão para estimação de idade em radiografias panorâmicas

Abstract

Age estimation is a relevant aspect of criminal and civil interest in the identification of corpses and refugees, in situations of individuals seeking asylum, work, school, access to social services, cases of rape, kidnapping and human trafficking. An age estimation study can be performed using dental analysis through tooth section or through imaging exams. Dental radiographic exams have a great advantage in assessing chronological age, as it is a non-destructive method. However, the interpretation of imaging exams is specific to human limitations. To minimize these views, studies propose the use of Artificial Intelligence (AI) to improve the prediction of chronological age. In this context, the present study aimed to estimate chronological age through a hybrid model using convolutional neural network (CNN) and regression models in panoramic images. For this, an observational, retrospective, cross-sectional study was carried out, which was approved by the Research Ethics Committee of the State University of Paraíba (opinion no. 6.700.058). A total of 2,546 panoramic radiographs of patients from the database of a private radiology clinic located in northeastern Brazil were selected. The images were acquired using a Carestream® CS 9000 3D panoramic scanner. Panoramic radiographs of patients over 18 years of age, of both sexes, with healthy and erupted lower first or second premolars were included. For image segmentation, the premolars present in each panoramic radiograph were delimited using the GNU Image Manipulation Program (GIMP® 2.10.36) software. The sample was divided into three data sets: model training (70% of the total sample); method validation (15% of the total sample); and test (15% of the total sample). A CNN model based on the MobileNetV2 architecture was developed to extract radiomic features. The task of estimating age was performed using a hybrid model that combined a CNN used to extract image attributes, and a regression model based on decision trees to study age using the attributes extracted from each image. The CNN extracted 576 attributes from each image. To analyze the performance of the proposed algorithm model, a widespread metric for evaluating systems for age assessment was applied, the Mean Absolute Percentage Error (MAPE). In addition to complementary analyzes, such as the Mean Absolute Error (MAE); Coefficient of determination or R2 Score; and Mean Squared Error (MSE). The verification of the CNN performance was done using the metrics module of the scikitlearn library (version 1.0.2). A total of 3,423 lower premolar teeth were analyzed, and the average age of the participants was 32.65 years (SD ± 9.71), ranging from 20 to 55 years of age. The performance of the proposed model, analyzing an entire database, showed a MAE of 6.52 years of age. However, when analyzing the sample by age group, the proposed model estimated a lower MAE in individuals aged 31 to 35 years, corresponding to an error of 0.29 years. In the analysis by sex, for images of male individuals in the age group of 31 to 35 years, the proposed model obtained an even higher accuracy with a MAE of only 0.16 years. Therefore, it is concluded that the proposed hybrid model is accurate and robust for patients aged 31 to 35 years. Therefore, it is expected that this model can assist in forensic investigations and estimate the age of people without official social identification.