Slice-to-volume registration for histology (H&E) ↔ tomographic volumes with a coarse-to-fine plane search, using MatchAnything (ROMA) inference, based on self-supervised learning DINOv2. This repository also includes a VALIS workflow to rebuild a 3D histology stack from serial sections prior to slice-to-volume co-registration.
-
MatchAnything (inference only): We import PL-LoFTR (ROMA) from your local MatchAnything clone to score plane candidates and compute a robust 2D affine transform between the resliced tomographic slab and the histology image.
-
VALIS (serial sections → 3D): We use VALIS to rigid/non-rigidly align serial histology sections and warp the original slides to a common frame, producing an aligned 3D stack that can be visualized or exported as OME-TIFF. VALIS reads WSIs via pyvips/OpenSlide that can be split in separate images if multiple sections were acquired on one slide.
