Deep Slice Interpolation for Reducing Through-Plane Anisotropy and Noise in Head CT

Researchers have developed a deep learning system that synthesizes intermediate CT slices to reduce through-plane anisotropy in head CT imaging, effectively halving spacing while simultaneously denoising outputs. The system outperforms classical interpolation and existing video frame interpolation methods, with MS-SSIM+L1 loss providing optimal performance across structural measures.

This research addresses a persistent technical limitation in medical imaging where head CT scans achieve sub-millimeter in-plane resolution but sacrifice through-plane spacing at 2-5 millimeters, creating significant anisotropy that compromises multiplanar reconstructions and volumetric measurements critical for clinical assessment. The deep learning approach tackles this fundamental constraint by synthesizing intermediate slices from adjacent axial images, effectively doubling spatial resolution in the through-plane dimension while delivering implicit denoising as a secondary benefit.

The systematic evaluation methodology distinguishes this work from incremental improvements. Rather than adopting a single loss function, researchers methodically compared pixel-wise losses (MSE, L1), structural-similarity approaches (SSIM, MS-SSIM), and hybrid combinations, documenting training instability patterns in SSIM-family losses and identifying partial remedies. This transparency regarding convergence challenges and batch-size dependent divergence provides practical guidance for practitioners implementing similar architectures.

The clinical validation extends beyond synthetic test sets through external deployment at Hospital Universitario Virgen del Rocío, where the model demonstrated generalization to out-of-distribution data while exhibiting the theoretically predicted denoising signature. This real-world validation strengthens claims that interpolation quality transcends training distribution boundaries.

For medical imaging workflows, this advancement directly improves diagnostic capability in stroke assessment, trauma evaluation, and surgical planning where hematoma volume estimation and three-dimensional visualization are clinically essential. The combined interpolation and denoising effects reduce downstream processing requirements while maintaining diagnostic fidelity. Healthcare IT departments and medical device manufacturers should monitor this approach's maturation toward clinical integration.

• →Deep learning system synthesizes intermediate CT slices, halving through-plane spacing while inherently denoising images from single inference pass
• →MS-SSIM+L1 hybrid loss provides strongest balanced performance, outperforming classical interpolation and pretrained video frame interpolation methods
• →Systematic documentation of SSIM loss training instability and batch-size dependent divergence offers practical implementation guidance
• →Real-world validation on external hospital data demonstrates generalization beyond training distribution with implicit denoising signature
• →Technology directly enables improved volumetric measurements and multiplanar reconstructions for stroke, trauma, and surgical planning applications