Optimizing µCT resolution in tarsal bones: A comparative study of super-resolution models for trabecular bone analysis

Abstract

Microcomputed tomography (µCT) is an essential tool for analyzing trabecular bone microarchitecture, yet its resolution is constrained by object size and acquisition time. To overcome these limitations, we implement a deep-learning-based super-resolution (SR) approach that enhances µCT image resolution while significantly reducing scan durations. Dry isolated tarsal bones (intermediate cuneiform) from 20 specimens were scanned using µCT at two resolutions, 80 µm voxel size (low resolution, LowRes) and 20 µm voxel size (high resolution, HiRes). Aligned LowRes and HiRes µCT data served as training data for SR reconstruction. In this study, we compare five SR models: 2D U-Net+ +, 3D SRCNN, 3D FSRCNN, 3D U-Net and a modified 3D U-Net model trained with a combined learned perceptual image patch similarity (LPIPS) and structural similarity (SSIM) loss function. The focus of this contribution is the application of these models based on real µCT data, rather than synthetically degraded images. Models were trained to learn volumetric representations for accurate restoration of trabecular bone microstructure. To assess SR image quality, we computed three image quality metrics (peak signal-to-noise ratio, SSIM and LPIPS) and evaluated bone morphometric parameters, i.e. average trabecular thickness (Tb.Th.) and bone volume fraction (BV/TV), across 95 regions of interest (ROI). RMSE was calculated for LowRes data and each SR model relative to HiRes data to quantify prediction accuracy. The results demonstrate that the 3D U-Net (LPIPS & SSIM) model achieves the highest reconstruction accuracy, yielding the lowest RMSE values (12.93 µm for Tb.Th. and 1.3 % for BV/TV), outperforming all other SR models in our evaluation. Compared to standard low-resolution µCT, our approach reduces scan time from 58 min to 7 min per sample while preserving trabecular morphology with high fidelity. These results demonstrate the effectiveness of perceptual loss-based SR to real µCT data for morphological analysis, ensuring accurate trabecular reconstruction and mitigating overestimation artifacts caused by LowRes imaging and partial volume effects. Integrating SR with real µCT scans offers a promising strategy to reduce scan time to improve throughput in bone imaging workflows. Future work will expand the training dataset to enhance model generalization across diverse bone structures and imaging conditions.