Classificação de cistos odontológicos em radiografias panorâmicas usando arquiteturas U-Net e Transfer Learning

Abstract

Odontogenic cysts are benign jaw lesions that may lead to structural changes and patient discomfort, making early detection clinically relevant. Panoramic radiography is widely used in Dentistry because it provides a global view of the maxillomandibular region and supports clinical decision-making. In this context, Artificial Intelligence (AI) methods can support specialists by reducing subjectivity and improving diagnostic efficiency. This work investigates the most effective methodology for classifying panoramic radiographs according to the presence or absence of odontogenic cysts using U-Net-based architectures adapted for binary classification. Two data sources were considered: a dataset built in collaboration with experts, comprising 604 panoramic radiographs, and the Tufts Dental Database, from which 24 images containing cyst regions of interest were selected. Regions of interest were annotated with LabelImg and automatically cropped based on the recorded coordinates, yielding 1219 cystic and 1212 non-cystic sub-images. To improve input quality, image pre-processing techniques were applied (BlackHat filter and Gamma Correction). Subsequently, U-Net v1, v2, and v3 variants were implemented and evaluated, with and without transfer learning using EfficientNetB0 as a pretrained backbone. Model validation was performed via stratified cross-validation (Stratified K-Fold) with 4 folds, using accuracy, sensitivity, specificity, and F1-score as evaluation metrics, aiming to identify the most robust configuration for the proposed task.