Comparison of different training datasets for AI superresolution in micro-CT imaging by means of Image Quality Assessment

Abstract

Reducing radiation exposure while maintaining image quality is a key challenge in X-Ray Computed Tomography (CT). One promising solution is the application of Artificial Intelligence (AI)-based Superresolution (SR) to enhance low-resolution CT scans. This thesis investigates how the choice of training dataset impacts the performance measured by Image Quality Assessment (IQA) of a 3D SR model applied to µCT imaging. Three types of training datasets were evaluated: (1) human bone specimens, (2) synthetic lattice structures, and (3) a combination of both. Additionally, the effect of the number of CT projections used during simulation (1440 vs. 360) was assessed. For each dataset, SR models were trained using a modified 3D U-Net architecture and evaluated with seven Full-Reference IQA metrics, namely PSNR, RMSE, SSIM, MS-SSIM, FSIM, VIFp, and VSI. A five-fold cross-validation ensured generalization and avoided data leakage. The findings indicate that dataset composition significantly affects SR model performance, with mixed datasets offering improved generalization across specimen types. Models trained on data simulated with fewer projections showed no significant difference in image quality to those with more projections but have the potential for radiation dose reduction.

Date of Award	2025
Original language	English
Awarding Institution	FH Gesundheitsberufe OÖ GmbH
Supervisor	Sascha Senck (Supervisor)