EDPS-SST: Enhanced Dynamic Path Stitching with Structural Similarity Thresholding for Large-Scale Medical Image Stitching Under Sparse Pixel Overlap

This paper studies large-scale medical image stitching under the challenging scenario of sparse pixel overlap, where there are few or no effective pixel overlaps between the images to be stitched (with an overlapping rate of only 3% to 6%). Existing stitching algorithms typically rely on rich pixel information for complex computations, whether it is feature extraction or deep learning methods. However, given sparse pixel overlap, very few feature points can be extracted for matching and hence the existing algorithms often fail in our scenario. To address this issue, we first propose a Structural Similarity Index (SSIM)-based algorithm, where the stitching is based on the structural similarity between the image pairs, and a novel SSIM thresholding strategy is designed to flexibly control the registration for stitching. Besides, when there are no pixel overlap, an Enhanced Dynamic Path Stitching (EDPS) method is proposed to improve stitching accuracy by considering a diverse range of path selection directions and referencing any images in the same row or column. Finally, some intelligent registration strategies are introduced to further improve the stitching accuracy and efficiency. Experimental results demonstrate the effectiveness of our method in terms of both stitching accuracy and efficiency, particularly in the scenario of sparse pixel overlap.