Deep learning for identifying lesions of the jaws in CBCT volumes: a pilot study

• Published in: Oral surgery, oral medicine, oral pathology and oral radiology Vol. 135; no. 2; pp. e51 - e52
• Main Authors: Huang, Yiing-Shiuan, Iakubovskii, Pavel, Lim, Li Zhen, Tyndall, Don
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.02.2023
• ISSN: 2212-4403; 2212-4411
• DOI: 10.1016/j.oooo.2022.10.019

This study aimed to train and validate a deep learning (DL) algorithm in the detection of radiolucent jaw lesions in CBCT volumes. CBCT volumes from 12 different scanners were acquired retrospectively from the UNC Adams School of Dentistry and Peking University School and Hospital of Stomatology. FOV ranged from 6 × 6 × 6 to 17 × 17 × 13 cm. CBCT volumes contained either zero or at least one intraosseous lesion. Volumes with no intraosseous lesions were included as control cases representing normal anatomy and were not annotated. Absence of lesions was verified by a board-certified oral and maxillofacial radiologist (OMR). Biopsy with a definitive diagnosis served as ground truth for lesion-positive cases. Annotations of lesion-positive volumes were made by a board-certified OMR and a second-year radiology resident using ITK-Snap. A total of 60 (50 positive, 10 control) CBCT volumes was presented to the DL software. Lesion-positive volumes contained an equal distribution of data from both institutions, and 40 of the lesion-positive volumes were used for initial training. Algorithm validation was performed using 10 controls and the 10 remaining lesion-positive volumes. The resulting model predictions and corresponding lesion annotations were visually compared. The DL algorithm demonstrated rudimentary pattern recognition and lesion detection. The predicted lesion size consistently showed a decreased superoinferior dimension. Visual inspection of model predictions showed a tendency for false positive artifacts (8/10 lesion-positive, 10/10 control). Almost all false positive artifacts presented as a distinctly linear shape and were easily recognizable. No lesion-like false positives were produced for the control volumes. Volumes which yielded false negative predictions (2/10 positive) contained lesions in interradicular areas. A DL algorithm trained on limited CBCT data showed preliminary ability to recognize patterns and detect intraosseous lesions in their correct locations. A larger training dataset is needed to improve quality of lesion detection. Human/Animal subjects were used and this study was approved by an institutional ethics panel.