Deep learning-driven super-resolution for cone-beam computed tomography: An ex vivo proof-of-concept study using...

This ex vivo proof-of-concept study aimed to develop deep learning (DL)-based super-resolution (SR) models to enhance simulated cone-beam computed tomography (CBCT) images.

Micro-computed tomography data from 51 extracted teeth were artificially degraded to simulate CBCT images. Three DL models, super-resolution convolutional neural network (SRCNN), local texture estimator (LTE), and Swin Transformer for image restoration (SwinIR), were compared with bicubic interpolation. Image quality was assessed using peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), learned perceptual image patch similarity (LPIPS), and deep image structure and texture similarity (DISTS). Three dentists evaluated sharpness and noise using a 5-point Likert scale. Eight observers assessed crack visibility in 47 images for LTE and bicubic interpolation using a 5-point Likert scale; scores were binarized using high and low thresholds.

All models significantly outperformed bicubic interpolation on objective metrics. SwinIR showed the highest PSNR (30.36 ± 2.66), whereas SRCNN achieved the highest SSIM (0.889 ± 0.073). LTE achieved the best LPIPS (0.253 ± 0.101) and DISTS (0.203 ± 0.049). Subjectively, LTE received the highest sharpness ratings (mean. 3.79 ± 0.47), whereas bicubic interpolation received the highest noise ratings (3.97 ± 1.43). LTE significantly improved crack visibility (odds ratio = 1.326, P = 0.006 for the low-threshold analysis; odds ratio = 1.310, P = 0.010 for the high-threshold analysis), with a higher pooled area under the curve (0.81 vs. 0.76, P = 0.063).

DL-based SR models can enhance simulated CBCT images, with LTE demonstrating superior perceptual sharpness and crack visibility.