Abstract
Modern spatial transcriptomics profiles the whole transcriptome
alongside matched H&E, yet remains expensive and, even on a
profiled slide, captures only part of the imaged tissue. This rich paired
data, together with vast archives of available H&E images, has motivated
methods that predict expression from H&E alone. Yet at the high
resolutions of current platforms this remains difficult, and recent work
suggests the limitation is structural rather than a shortcoming of any
one model. At Visium HD’s 8 μm resolution a bin integrates so few transcripts
that per-gene counts are dominated by dropout, and contrastive
image-expression alignment degrades direct gene prediction further.
We argue that predicting counts directly is the wrong target. This reflects
how expression is used downstream, where cell types, states, and
spatial programs are derived from a denoised, batch-corrected latent.
CIVET predicts that latent instead. A frozen H-Optimus-1 foundation
model, adapted with low-rank adapters, is aligned to each bin’s DrVI
latent under a contrastive objective, with a CoCa-style gene-panel decoder
anchoring the embedding to marker expression, and imputes uncaptured
tissue at 8 μm by soft-kNN retrieval. The same retrieval yields
a per-prediction confidence at no cost, indicating where morphology is
out-of-distribution. Under leave-one-sample-out validation on five public
Visium HD colorectal sections, CIVET recovers the latent at a median
active-dimension Pearson of 0.55±0.04, against 0.28 for a frozenbackbone
baseline; decoded back to genes, it further outperforms methods
that predict expression directly. Treating this as in-distribution imputation
rather than cross-cohort generalisation, CIVET extends each
section’s expression map roughly 4.7x into contiguous unprofiled tissue.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/COMPAYL_045.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: https://papers.miccai.org/miccai-2026-sat/supp/COMPAYL_045_supp.pdf
Link to Open Review
BibTex
@InProceedings{TreTim_Beyond_MICCAISAT2026,
author = { Treis, Tim AND Lubeck, Eric AND Theis, Fabian J.},
title = { { Beyond the capture area: Imputation of spatial gene expression from H&E } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17251},
month = {pending},
page = {pending}
}