Evaluation of deep learning for detecting intraosseous jaw lesions in cone-beam computed tomography volumes

Yiing-Shiuan Huang; Pavel Iakubovskii; Li Zhen Lim; André Mol; Donald A Tyndall

doi:10.1016/j.oooo.2023.09.011

Evaluation of deep learning for detecting intraosseous jaw lesions in cone-beam computed tomography volumes

Oral Surg Oral Med Oral Pathol Oral Radiol. 2023 Oct 16:S2212-4403(23)00690-9. doi: 10.1016/j.oooo.2023.09.011. Online ahead of print.

Authors

Yiing-Shiuan Huang¹, Pavel Iakubovskii², Li Zhen Lim³, André Mol⁴, Donald A Tyndall⁴

Affiliations

¹ Oral and Maxillofacial Radiology, Adams School of Dentistry, University of North Carolina, Chapel Hill, NC, USA. Electronic address: yiingshiuanhuang@gmail.com.
² Denti.AI Technology Inc., Toronto, Ontario, Canada.
³ Oral and Maxillofacial Radiology, Adams School of Dentistry, University of North Carolina, Chapel Hill, NC, USA; Discipline of Oral and Maxillofacial Surgery, Faculty of Dentistry, National University of Singapore, Singapore.
⁴ Oral and Maxillofacial Radiology, Adams School of Dentistry, University of North Carolina, Chapel Hill, NC, USA.

PMID: 38155015
DOI: 10.1016/j.oooo.2023.09.011

Abstract

Objective: The study aim was to develop and assess the performance of a deep learning (DL) algorithm in the detection of radiolucent intraosseous jaw lesions in cone-beam computed tomography (CBCT) volumes.

Study design: A total of 290 CBCT volumes from more than 12 different scanners were acquired. Fields of view ranged from 6 × 6 × 6 cm to 18 × 18 × 16 cm. CBCT volumes contained either zero or at least one biopsy-confirmed intraosseous lesion. 80 volumes with no intraosseous lesions were included as controls and were not annotated. 210 volumes with intraosseous lesions were manually annotated using ITK-Snap 3.8.0. 150 volumes (10 control, 140 positive) were presented to the DL software for training. Validation was performed using 60 volumes (30 control, 30 positive). Testing was performed using the remaining 80 volumes (40 control, 40 positive).

Results: The DL algorithm obtained an adjusted sensitivity by case, specificity by case, positive predictive value by case, and negative predictive value by case of 0.975, 0.825, 0.848, and 0.971, respectively.

Conclusions: A DL algorithm showed moderate success at lesion detection in their correct locations, as well as recognition of lesion shape and extent. This study demonstrated the potential of DL methods for intraosseous lesion detection in CBCT volumes.