Deep CT Slice Interpolation Halves Anisotropy, Cuts Noise in One Pass

A team led by Luis Cortés Ferre from Hospital Universitario Virgen del Rocío and other institutions has published a paper on arXiv proposing a deep learning system for head CT slice interpolation. Head CT scans typically have sub-millimeter in-plane resolution but 2-5 mm through-plane spacing, causing significant anisotropy that degrades multiplanar reconstructions and volumetric measurements like hematoma volume. The new model takes pairs of neighboring axial slices and synthesizes intermediate slices, effectively halving the through-plane spacing. Remarkably, the same inference pass also produces inherently denoised outputs, offering two benefits from one operation.

To build a reliable system, the team systematically evaluated various loss functions: pixel-wise losses (MSE, L1), structural similarity losses (SSIM, MS-SSIM), and hybrid combinations. On a held-out test set, all converged models outperformed classical interpolation baselines and pretrained video frame interpolation methods (RIFE, FILM) on all structural measures. The MS-SSIM+L1 hybrid offered the strongest balanced profile. They also documented training instability with SSIM-family losses and identified partial remedies. As an illustration, the system was applied to an out-of-distribution head CT series from Hospital Universitario Virgen del Rocío, supporting that interpolation quality and implicit denoising generalize beyond the training distribution.