Utilização do aprendizado de máquina no diagnóstico de lesões periapicais crônicas por meio de imagens tomográficas

Abstract

Periapical granuloma (PG) and radicular cyst (RC) are chronic inflammatory periapical lesions that arise at periapical region, with high prevalence rates and similarity between their clinical and imaginological features. To performed a differential diagnosis of PG and RC on Cone-Beam Computed Tomography (CBCT) still remains difficult. In this context, machine learning have been used to improve the accuracy of image diagnosis. Objective: To compare the accuracy of different classifiers, based on machine learning on tomographic images, of cases of PG and RC histopathologically diagnosed. Materials and Methods: This study was observational, transversal and analytical images of histopathologically diagnosed GPs and CRs. Sample consisted of 14 cases of PGs and 9 cases of RCs. CBCT images were obtained using the CS 9000 3D system (Carestream Dental). Coronal reconstructions of the CT scans were evaluated in 05 different slices in CS3D Imaging Software. In GIMP software, version 2.10, the Region of Interest (RoI) was segmented and exported in TIFF format image with 32 bits. After segmentation, 6 texture attributes (contrast, dissimilarity, correlation, energy, homogeneity, and second angular momentum) were extracted from the RoI through of co-occurrence matrices. After extraction of these attributes, principal component analysis was used to perform dimensionality reduction and select the most significant attributes. A machine learning classifier algorithm (K-means) was used to calculate the accuracy, sensitivity, and specificity of method. Data were analyzed using Hopkins statistics to identify the presence of clusters. Results: Regarding the clinical findings, 72.0% (n=8) of the PG were woman, and 100% of RC were man. The mean age was 47.33 years for RCs and 35,50 years for PGs. For machine learning, 105 images were reviewed and 77.14% accuracy, 47.06% sensitivity and 91.55% specificity were observed in the grouping of lesions. Conclusion: It is possible suggest that use of Artificial Intelligence shows promising results for the differentiation of chronic periapical lesions through tomographic images.