DEMIS: Electron Microscopy Image Stitching using Deep Learning Features and Global Optimisation

Accurate stitching of overlapping image tiles is essential for reconstructing
large-scale Electron Microscopy (EM) images during Whole Slide Imaging. Current
stitching approaches rely on handcrafted features and translation-only global
alignment based on Minimum Spanning Tree (MST) construction. This results in
suboptimal global alignment since it neglects rotational errors and works only
with transformations estimated from pairwise feature matches, discarding valuable
information tied to individual features. Moreover, handcrafted features may have
trouble with repetitive textures. Motivated by the limitations of current methods
and recent advancements in deep learning, we propose DEMIS, a novel EM image
stitching method. DEMIS uses Local Feature TRansformer (LoFTR) for image
matching, and optimises translational and rotational parameters directly at the
level of individual features. For evaluation and training, we create EM424,
a synthetic dataset generated by splitting high-resolution EM images into arrays
of overlapping image tiles. Furthermore, to enable evaluation on unannotated
real-world data, we design a no-reference stitching quality metric based on
optical flow. Experiments that use the new metric show that DEMIS can improve the
average results from 32.11 to 2.28 compared to current stitching techniques (a
1408% improvement).