Detecção automática de cistos odontológicos em imagens de radiografia panorâmica

Abstract

Odontogenic cysts are pathological cavities lined with epithelium and have an intraosseous occurrence in the mandibular region. They present erosion and expansive growth, usually with asymptomatic characteristics, when they expand considerably causing disorder and pressing other structures causing pain to the patient. Due to this expansive capacity, it is necessary to identify these pathologies in advance to prevent their spreading along the facial sinuses and endangering the patient’s life. The best way to identify these pathologies is through the diagnosis by dental radiographs, which represent a major advance for the dental area. Even with the progress offered by these images, identifying cysts is not a simple task, their interpretation is challenging even for experienced specialists. In addition, because images are subjectively analyzed, several factors can interfere, such as limitation of the human vision system, fatigue and stress. Thus, considering the need for diagnostic support systems, this study proposes a methodology for identifying cystic regions in panoramic radiography images. For this, investigations of classical techniques of digital image processing along the steps are performed, such as mathematical morphology, Otsu multilevel, thresholding, active contours and others. Analyzes are also performed with the Local Binary Pattern and Haralick texture descriptors, used for classification with the methods for learning artificial neural network (RNA), Support Vector Machine (SVM) and Extreme Learning Machine (ELM) machines. Inaddition, the methodology also implements a Convolutional Neural Network (CNN) confronting a classical approach to computer vision with a modern model of deep learning. The results of the methodology are compared with ground truth and obtained as the best approach to CNN with an average Dice of 0.86. Other results of the investigations conclude that segmentation in the internal region of the mandible gives the best Otsu multilevel, texture analysis obtains better metrics through the Haralick descriptor with ELM and RNA. Finally, segmentation studies from the pathological region resulted in the active contours technique with the morphological filter as the best combination for the final segmentation.