Framework de aprendizagem de máquina para reconhecimento automático de cistos odontológicos em imagens de radiografia panorâmica

Abstract

Odontogenic cysts are part of oral and maxillofacial pathology, commonly seen in dental practice. They can cause anything from painless expansion of the bone in the area involved, to facial asymmetry, pain and swelling. To diagnose them, dentists use panoramic radiography (PAN), which provides a global view of the teeth, supporting bone tissues and adjacent anatomy. In this case, the interpretation of the PAN helps to reduce surgical risks in the posterior region of the mandible in anesthesia, implants, tooth extractions and other procedures. Thus, it can be assumed that erroneous PAN analyses lead to delays in correctly diagnosing a present oral health problem or preventing one in the future. This can be mitigated with more effective analytical tools that learn from past data, such as exams, reports and prescriptions. In this case, Artificial Intelligence (AI) algorithms can help dentists get a second opinion on their diagnoses and also help train future dentists. In this work, a reproducible PAN processing protocol is developed with AI software to automate the recognition of dental cysts (commonly caused by third molars, known as “wisdom teeth”) in PANs. First, databases of public and private PANs were collected. The images were then pre-processed to highlight their details using a filter, and a binary mask was applied to remove information irrelevant to the problem. Then, they were reduced in resolution so that training the Machine Learning (ML) model would use less computational resources. Consequently, the coordinates of the bounding boxes that locate each dental cyst also had to be corrected to be consistent with the new resolution of the images. Following that, various data augmentation techniques were applied to expand the number of samples and make the ML model more generalizable to unknown cases. After initial tests with ML architectures, the best performance with a pre-trained model was seen in MobileNetV3Large, which was modified by adding extra layers of neurons, activation functions and feature map refinement. In the end, in addition to the labeling of the partially unpublished PAN dataset, there was an improvement in the Mean Squared Error (MSE) metric compared to the standard version of MobileNetV3Large. This improvement in MSE resulted in the observed and logged values of the customized Bounding Box Accuracy and Intersection over Union metrics.