Diffusion Refinement for Boundary-Consistent Veterinary X-ray Reconstruction

In AI-based analysis pipelines, large medical images are often restored in patches for high-resolution processing due to memory limitations. Patch-wise restoration approaches cause brightness and texture stitching artifacts at patch boundaries, degrading structural clarity. In this study, we propose a method to remove artificial lines when generating high-resolution X-ray images from patches. We present a selective refinement pipeline focusing on patch boundaries. Using pre-restored high-resolution patches, we apply a soft boundary mask and perform noise-based reconstruction only in boundary areas. This process helps blend patch edges smoothly without extra post-processing. Experimental results show that the proposed method improves both structural and visual quality. Artificial seams are reduced, and the shapes of bones and tissues appear more natural and continuous. This approach offers a practical way to improve patch-based medical image restoration by making boundaries clearer and more consistent.