Parametric kernels for artifact mitigation in patch-based image aggregation using generative models

Recent advancements in artificial intelligence applications have highlighted the effectiveness of generative models for domain transfer, image enhancement and simulation. However, when applied to large-scale gigapixel images, the use of traditional patch-based image aggregation methods introduces checkerboard or blocking artifacts, which compromises image quality and fidelity. In this paper, we propose a parametric kernel that is specifically designed to target the underlying grid structure to mitigate these artifacts. With the use of adjustable zero-padding and linear-padding parameters, our kernel provides fine control over the fusion process by combining a central area of constant weights with border regions of gradually decreasing weights. The proposed method was validated using three medical imaging modalities (digital pathology, fluorescence microscopy and ultrasound imaging) for different generative model tasks. The results showed statistically significant improvements (p < 0.0001) in artifact removal when compared to state-of-the-art methods. Quantitative analysis revealed improvements in fusion quality measures for digital pathology, fluorescence microscopy and ultrasound imaging of 7.1%, 27.4% and 20.0%, respectively. Additionally, expert evaluators confirmed superior visual quality and reduced artifacts in blind assessments of reconstructed images, with our method achieving significantly higher scores across all modalities. Our method is versatile, compatible with various generative models and can be easily adjusted by modifying kernel parameters. This kernel-based approach significantly advances the quality of synthesized medical images, directly supporting more reliable clinical assessment and automated analysis.