Deep Learning-Based Super-Resolution Applied to Dental Computed Tomography

The resolution of dental computed tomography (CT) images is limited by detector geometry, sensitivity, patient movement, the reconstruction technique and the need to minimize radiation dose. Recently, the use of convolutional neural network (CNN) architectures has shown promise as a resolution enhancement method. In the current work, two CNN architectures – a subpixel network and the so called U-net – have been considered for the resolution enhancement of 2D cone-beam CT image slices of ex vivo teeth. To do so, a training set of 5680 cross-sectional slices of 13 teeth and a test set of 1824 slices of 4 structurally different teeth were used. Two existing reconstruction-based super-resolution methods using l2-norm and total variation regularization were used for comparison. The results were evaluated with different metrics (peak signal-to-noise ratio, structure similarity index, and other objective measures estimating human perception) and subsequent image-segmentation-based analysis. In the evaluation, micro-CT images were used as ground truth. The results suggest the superiority of the proposed CNN-based approaches over reconstruction-based methods in the case of dental CT images, allowing better detection of medically salient features such as the size, shape, or curvature of the root canal.