Impact of Generative AI-Enhanced Low-Dose Cone-Beam Computed Tomography on Diagnosis and Treatment Planning for Impacted Mandibular Third Molars.
Authors
Affiliations (6)
Affiliations (6)
- Department of Diagnostic Radiology, Li Ka Shing Faculty of Medicine, The University of Hong Kong, Hong Kong SAR, China.
- Oral and Maxillofacial Radiology, Applied Oral Sciences and Community Dental Care, Faculty of Dentistry, The University of Hong Kong, Hong Kong SAR, China. Electronic address: [email protected].
- Oral and Maxillofacial Surgery, Faculty of Dentistry, The University of Hong Kong, Hong Kong SAR, China.
- Oral and Maxillofacial Radiology, Applied Oral Sciences and Community Dental Care, Faculty of Dentistry, The University of Hong Kong, Hong Kong SAR, China.
- Restorative Dental Sciences, Faculty of Dentistry, The University of Hong Kong, Hong Kong SAR, China.
- Department of Diagnostic Radiology, Li Ka Shing Faculty of Medicine, The University of Hong Kong, Hong Kong SAR, China. Electronic address: [email protected].
Abstract
To evaluate whether generative artificial intelligence (Gen-AI) could significantly enhance the visibility of the mandibular canal (MC) and the periodontal ligament (PL) of mandibular third molars (M3Ms) on low-dose cone-beam computed tomography (CBCT) images compared to standard-dose images, and to assess its impact on clinical decision-making compared to standard- and low-dose CBCT. A total of 302 CBCT scans with 151 paired from 90 patients with impacted M3Ms were acquired using one standard-dose (333 mGy × cm2) and various low-dose (78-131 mGy × cm2) protocols. Gen-AI models (Pix2Pix, CycleGAN, and diffusion models) were trained using paired standard- and low-dose CBCT images, with the CycleGAN-based model demonstrating superior performance. Quantitative image quality was assessed using the structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), mean absolute error (MAE), and root mean square error (RMSE). Three blinded clinicians, a general practitioner (GP), oral-maxillofacial surgeon (OMFS) and oral-maxillofacial radiologist (OMFR), evaluated the MC and PL visibility, M3M-MC proximity, root morphology, adjacent molar status, surgical approach, and referral decisions. Pairwise comparisons were performed using Wilcoxon signed rank test. The quality of Gen-AI-enhanced low-dose CBCT was significantly improved, achieving higher PSNR, lower MAE, and lower RMAE compared to the original low-dose CBCT (all P < .001), and maintained excellent anatomical fidelity with an SSIM of 0.97 compared to standard-dose CBCT. Gen-AI-enhanced low-dose images showed significantly higher MC visibility for all clinicians and higher PL visibility for both the GP and OMFS compared to low-dose images. No significant differences were observed for other variables. Gen-AI-enhanced low-dose CBCT images significantly improved the visibility of the MC and PL for M3M evaluation. Compared to the original CBCTs, these AI-enhanced low-dose images did not significantly affect risk assessments, treatment strategies, or patient management decisions, and were largely indistinguishable from original images.