Avaliação da relação entre a idade do paciente e a estrutura óssea trabecular da mandíbula através da análise de imagens tomográficas dentárias cone beam por meio de uma rede neural convolucional

Abstract

Osteoporosis is a condition that affects bone mineral density in individuals, making it difficult to insert dental implants into the jaw bone of patients. It mainly affects women and is diagnosed by DXA (Dual Energy X-Ray), considered the gold standard in the diagnosis of osteoporosis. However, the DXA test is not easily accessible by a large part of the Brazilian population. The Jaw System Age Group X software, created in this research, aims to bring patients closer to the assessment of the trabecular bone structure of their mandibles; the software itself will not diagnose osteoporosis but will provide the dentist with information to refer the patient to a doctor who will eventually request the DXA exam, optimizing the use of SUS equipment. The software uses cone beam tomography in its analysis, performs image segmentation processes and a neural network to classify the trabecula of the mandible bone, in the coronal anatomical plane, into age classes, respecting the patient's sex. The software user manipulates the tomographic images, choosing the slices that will be segmented and analyzed by the neural network. The research used 137 cone beam CT scans, 52 of male patients and 85 of female patients, provided by the UNIOESTE Dental Clinic. The database to train the convolutional neural network consisted of 1389 samples from the mandible from female patients and 633 from male patients. The neural network is composed of 3 dense layers with 100 neurons each and Relu activation function, with weights updated through the Adam algorithm, using MaxPooling in each convolution; for image classification, it uses a dense layer with 100 neurons, with the softmax activation function to classify the samples into possible classes. The neural network training accuracy was 99% for males and 96% for females, with the area under the ROC curve (AUC) equal to 0.81 and 0.97 for the respective sexes. The accuracy obtained in the CNN validation was 70,66% for males and 90,89% for females. Finally, a supervised test was performed, using 5 cuts from 10 exams of the test set for each sex; considering the average result, approximately 100% accuracy was obtained for both sexes. Thus, it is concluded that the proposed model can classify the samples in the proposed age groups and proved to be robust and solid in the tests. Therefore, this tool can help professionals who use it, being an alternative tool, with low costs, for bone density analysis, and can serve as a filter for diagnostic tests, such as DXA