Deep µStitch: an unsupervised sliding window-based deep global microscopy image stitching framework

Microscopy plays a pivotal role in biological and medical research, enabling detailed observation of specimens. Nevertheless, the inherent constraint of a micro-objective's small field of view (FoV) restricts conventional microscopy to observing exceedingly small regions. While techniques such as lensless holography, ptychography, and multi-camera array-based gigapixel imaging significantly expand the FoV, each approach presents its own challenges: lensless holography struggles with limited spatial resolution, whereas ptychography and multi-camera array-based gigapixel imaging necessitate prolonged image reconstruction time and intricate optical setups. To balance spatial resolution, speed, and cost, FoV scanning and stitching-based whole slide imaging emerge as an optimal solution. Yet, numerous FoV stitching methods suffer from pronounced artifacts and underperform when dealing with low-contrast images. To mitigate these limitations, we devised Deep µStitch, which is an unsupervised sliding window-based deep global microscopy image stitching framework and has been effectively deployed in both bright-field and fluorescence microscopy applications. Furthermore, we conducted a comparative analysis of Deep µStitch against prevalent microscopy stitching tools, notably the microscopy image stitching tool (MIST), grid stitching, and scale-invariant feature transform (SIFT). These findings suggest that our proposed Deep µStitch achieves high FoV stitching quality, boasting nearly the highest peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM), while minimizing visible stitching artifacts. Given its outstanding FoV stitching capabilities, Deep µStitch presents a promising approach for whole slide imaging.