Odonto Vision: sistema cad para segmentação automática de dentes em radiografias panorâmicas.

Abstract

Image segmentation is fundamental in automatic analysis applications, as it allows for the precise identification and delineation of relevant structures. In the dental context, this task is especially important in panoramic radiographs, which are widely used to provide a comprehensive view of the dental and bone structures. However, manual interpretation of these images is timeconsuming, subject to inter-professional variability, and affected by noise and low contrast. Recent advances in deep learning, particularly Convolutional Neural Networks (CNNs), have enabled more accurate segmentations through the automatic extraction of relevant features. Thus, this study aims to develop a CAD system to perform automatic tooth segmentation from panoramic radiographs, based on a comparative analysis of the performance of five U-Net–based architectures. The study was conducted using the Children’s Dental Panoramic Radiographs Dataset, in which the images underwent a preprocessing stage, and the models were trained using K-Fold cross-validation combined with Grid Search for hyperparameter optimization. The results of the external validation stage, conducted on two subsets of the overall dataset, showed that in the first subset, Dataset and code, the W-Net achieved the best performance, with Sensitivity 0.809, E-MEASURE 0.925, IOU 0.785, and DSC 0.879, indicating a higher capacity for segmenting dental regions. In the second subset, Panoramic Radiography Database, the Attention U-Net stood out, reaching Sensitivity 0.965, E-MEASURE 0.975, IOU 0.858, and DSC 0.924, showing higher accuracy, preservation of tooth shape, and lower fragmentation of structures. The Wilcoxon test indicated that U-Net++ and U-Net 3+ showed significant differences compared to the classical U-Net, while W-Net and Attention U-Net had equivalent performance. Finally, considering the results of both quantitative and qualitative analyses of the segmented masks, as well as the inference time required to generate the segmentations, the Attention U-Net was selected as the final architecture to be integrated into the CAD system.