TiSAM: fusing spatial and histopathology foundation models for whole-slide image segmentation

Ding, Z.; Dequiedt, L.; Forjaz, M.; Matos-Romero, V.; Shen, Y.; Benducci, T.; Forjaz, A.; Fei, X.; Zhu, Y.; Nwoga, P.; Dhana, W.; Pascalin, J.; Lee, S. S.; Garza, L.; Lee, S. H.; Pierce, E. J.; Canto, M. I.; Morton, J. P.; Cihakova, D.; Wirtz, D.; Data Science TeamLab,; Kiemen, A. L.

doi:10.64898/2026.10.16.759688

Home/Archive/bioRxiv

bioRxiv · 10.64898/2026.10.16.759688

TiSAM: fusing spatial and histopathology foundation models for whole-slide image segmentation

View original source ↗Export metadata ↓

Abstract

Semantic segmentation of whole-slide images (WSIs) quantifies tissue composition and architecture, but existing models require extensive annotation and lose accuracy as the number of tissue classes increases or on heterogeneous, cross-scanner datasets. Here, we present the Tissue Segment Anything Model (TiSAM), which fuses SAM3 spatial features with UNI2-h pathology features within a multi-scale decoder. Because public benchmarks contain few tissue classes, we release MegaMouse, comprising 36,833 annotations of 52 tissue classes from 12 mouse organs, and an 11-class human pancreas dataset. TiSAM achieved the highest overall performance across four public datasets compared to existing convolutional neural network and vision transformer models, maintained high accuracy on MegaMouse, and maintained high accuracy in the human pancreas dataset when trained on as few as 5 annotated WSIs where previous integrations of SAM with pathology foundation models did not. TiSAM segmentations were also more consistent across eight slide scanners compared to other models. These results establish TiSAM as a general and highly adaptable framework for tissue segmentation, enabling studies containing many tissue classes, limited annotations, or images from multiple scanners.

Explore related subjects

bioinformatics Preprint

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Ding, Z., Dequiedt, L., Forjaz, M., Matos-Romero, V., Shen, Y., Benducci, T., Forjaz, A., Fei, X., Zhu, Y., Nwoga, P., Dhana, W., Pascalin, J., Lee, S. S., Garza, L., Lee, S. H., Pierce, E. J., Canto, M. I., Morton, J. P., Cihakova, D., Wirtz, D., Data Science TeamLab,, Kiemen, A. L.. 2026-10-19. TiSAM: fusing spatial and histopathology foundation models for whole-slide image segmentation. https://doi.org/10.64898/2026.10.16.759688

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

MINT Block: multivariate integration across independent studies to identify multi-omics signatures

Integrative multi-omics analyses can identify molecular signatures spanning multiple layers of a biological system. However, signatures derived from datasets with small sample sizes may not generalise. Combining data from multiple independent studies can increase the available sample size, but may also introduce unwanted variation arising from study-specific differences. Existing approaches typically address this variation sequentially by correcting the data before downstream model fitting, which can inadvertently distort the biological signals of interest. We introduce MINT Block, a framework that integrates multi-omics data across independent studies while simultaneously accounting for study-specific unwanted variation. MINT Block supports multivariate regression and classification analyses, enabling the identification of generalisable multi-omics signatures that capture relationships between different types of molecules and predict sample phenotypes. We evaluated MINT Block using bulk and single-cell multi-omics datasets from studies of human disease and marine biology. Across these applications, we showed that MINT Block identified stable, interpretable molecular signatures and improves prediction accuracy compared with existing sequential approaches.

bioinformatics↗

bioRxiv2026

BiHC: Protein domain segmentation by bidirectional hierarchical clustering using Voronoi tessellation and physical self-containment

Protein domains are commonly regarded as structurally separated, functionally relevant, and evolutionarily separable units. Although classification systems may assign different domain boundaries to the same structure, structural separation remains a shared criterion, and many domain parsers assess it through inter-residue distances, contact numbers, or contact densities. However, the Voronoi contact area, which is widely used to study protein-protein interactions and predict contact sites, has seen limited use in domain parsing. There are two potential challenges to this problem. First, unlike in protein-protein interaction, partitions are not known in advance and must be identified before their interface can be evaluated. Second, the contact area depends on fragment size and internal packing, making it difficult to compare candidate partitions. Here, we introduce BiHC (Bidirectional Hierarchical Clustering), a domain parser based on all-heavy-atom Voronoi contact areas. By normalizing the contact area for fragment size and internal packing, and incorporating residue hydropathy, we define a pairwise self-containment score F. BiHC applies F in two independent hierarchical searches: top-down recursive splitting until no admissible division remains, and bottom-up merging of secondary-structure elements until the entire structure becomes a single fragment. Intermediate partitions from both searches are ranked by a global partition quality score G, derived on the same principle. On two CATH-ECOD consensus benchmarks, under balanced evaluation, BiHC achieved the strongest performance among all tested criterion-based parsers and was comparable to Chainsaw, the best-performing deep learning method tested. These results show that the Voronoi-based criterion can yield competitive domain parsing.

bioinformatics↗

bioRxiv2026

GOLDEN: Gene expression and allele-informed decomposition for genetic heterogeneity in spatial transcriptomic data

Motivation Spatial transcriptomics (ST) enables exploration of cellular interactions, including immune responses. However, current methods rely on pathologists, with limitations in addressing genetic heterogeneity, where niches harboring nonclonal cells, including placental or allograft tissues, are present. Deciphering these interindividual mixtures remains challenging. Results For deciphering interindividual heterogeneity in ST, we developed GOLDEN (Gene expressiOn and aLlele-informed DEcomposition for geNetic heterogeneity), which involves an allele-aware multitask network that encodes spot-level expressions into shared latent representations. Incorporating allele features, dual heads classify spot identity (target, nontarget, or mixed) and estimate target fractions. We utilized synthetic ST data with spot-level ground truth using semiallogeneic single-cell profiles. For these synthetic benchmarks, GOLDEN achieved reliable deconvolution, with high correlation (Pearson's r = 0.977) and low error (mean absolute error (MAE) = 0.045) in fraction estimation and consistent predictions across real-world STs. These predictions were genetically validated. Using GOLDEN, resolving spot-level genetic heterogeneity is feasible. Availability and implementation GOLDEN and the simulation code for ST are available at https://github.com/othertics/GOLDEN.

bioinformatics↗