Impact of Gaussian Feathering on Diagnostic Metrics in Tile-Based Micro-CT Sinogram Infilling

Tile-based generative AI methods for micro-CT sinogram infilling require merging overlapping tiles to reconstruct full images. Gaussian feathering, the standard blending approach, produces visually seamless results but its effect on diagnostic image quality metrics has not been characterized. This study quantifies how Gaussian feathering affects noise power spectrum (NPS), modulation transfer function (MTF), noise equivalent quanta (NEQ), structural similarity index (SSIM), and peak signal-to-noise ratio (PSNR) compared to nearest priority blending, also known as Voronoi partition blending. We mathematically derived the variance reduction mechanism in Gaussian blending and experimentally compared both methods using DeepFill v2 infilled sinograms from a micro-CT quality assurance phantom, reconstructed from 50% undersampled data. Gaussian feathering reduced variance by up to 19.6% at overlap centers, causing NPS reduction of up to 13.8% at low spatial frequencies and NEQ inflation of up to 38.5%. MTF showed mixed effects with improvements at low frequencies but reductions at high frequencies. SSIM and PSNR showed statistically significant differences: sinogram PSNR differed by 0.05 dB (p = 0.003) and reconstruction SSIM by 0.009 (p = 0.036), both with small effect sizes. Gaussian feathering distorts diagnostic metrics through a variance reduction mechanism, while standard fidelity metrics detect only subtle changes. Nearest priority blending should be preferred when diagnostic metrics are used to validate infilling methods.