Deep Learning‐Based Detection of Simulated Root Resorption in Scenarios Involving Image‐Degrading Artifacts: An in Vitro Study
Orlando Aguirre Guedes, Letícia Junqueira de Pádua Sesti Gomes Moussa, Lucas Rodrigues de Araújo Estrela, Raoni Florentino da Silva Teixeira, Cyntia Rodrigues de Araújo Estrela, Raul Vitor Arantes Monteiro, Gordon Lai, Carlos EstrelaABSTRACT
Background/Aim
This study developed a deep learning‐based convolutional neural network (CNN) model for detecting external root resorption (ERR) in periapical radiographs and cone‐beam computed tomography (CBCT) scans, particularly in the presence of image‐degrading artifacts.
Material and Methods
A total of 480 bovine incisors were allocated into four experimental conditions ( n = 120) according to the presence and absence of root canal treatment and ERR. Resorption defects were made 5 mm from the root apex using a round diamond bur. All specimens were imaged using periapical radiography and CBCT under standardized acquisition protocols. CBCT images were processed using a post‐processing CBCT software, with and without the Blooming Artifact Reduction (BAR 1) algorithm, and with a volumetric rendering reconstruction tool. A pre‐trained AlexNet CNN was adapted using transfer learning for four‐class image classification. The CNN performance was evaluated using standard classification metrics, including overall accuracy, precision, recall, and F1‐score.
Results
CNN performance varied across imaging modalities. Periapical radiography yielded perfect classification (100% accuracy). High accuracy was also observed for CBCT without BAR 1 (95.83%) and CBCT with BAR 1 (97.92%). CBCT with three‐dimensional reconstruction showed reduced performance (73.96%), particularly in endodontically treated teeth without root resorption.
Conclusions
Deep learning‐based CNN model demonstrated high diagnostic performance for detecting ERR, strongly influenced by image acquisition and processing protocols.