Craniofacial CBCT: Addressing volume-resolution dilemma using generative artificial intelligence.
Authors
Affiliations (5)
Affiliations (5)
- Department of Bioengineering, McGill University, Montreal, Canada. Electronic address: [email protected].
- Department of Anthropology, University of Western Ontario, London, Canada.
- Comet Technologies Canada Inc., Montreal, Canada.
- Faculty of Dental Medicine and Oral Health Sciences, McGill University, Montreal, Canada.
- Department of Bioengineering, McGill University, Montreal, Canada; Faculty of Dental Medicine and Oral Health Sciences, McGill University, Montreal, Canada; Department of Anatomy and Cell Biology, McGill University, Montreal, Canada.
Abstract
In large field-of-view tomographic images, fine structures such as bone texture, porosity or root canals are poorly resolved. We developed a Generative-Adversarial-Network (GAN)-based super-resolution method for craniofacial cone-beam CT (CBCT) that recovers sharp edges while avoiding hallucinations. CBCT scans of 1 human skull, 4 cadaveric human heads, and 6 sheep heads were acquired at 270 μm and 135 μm voxel, and a GAN (Unet-Structure-Preserving Super-Resolution [UNetSPSR]) was designed to recover fine image details in the low-resolution images. Performance was benchmarked against state-of-the-art methods using peak-signal-to-noise-ratio (PSNR) and learned-perceptual-image-patch-similarity (LPIPS), and evaluated in segmentation of trabecular bone and root canals. Independent archival clinical CBCT scans from five patients and an external public dataset (n = 50) were used for generalization testing. UNetSPSR achieved the highest PSNR (+0.66 on seen and +0.11 on unseen data relative to the runner-up) and the lowest LPIPS (-0.01), indicating superior recovery of fine image details. The comparable performance between seen and unseen internal test sets, demonstrates strong generalization within the target data distribution. On independent clinical and external datasets, the network enhanced fine structures without introducing visible artifacts, despite heterogeneous acquisition parameters, although the absence of high-resolution reference standards precluded definitive validation. UNetSPSR substantially reduced overestimation of trabecular morphology, decreasing trabecular thickness bias from 61% to 11%, and significantly improved microarchitectural measures on external data (p < 0.05). Root canal segmentations more closely resembled high-resolution references, with improved quantitative metrics. UNetSPSR enables structure-preserving super-resolution of craniofacial CBCT, improving depiction of fine anatomical features. While results demonstrate internal generalization, further external validation is warranted.