SpaTIE

The reporting vocabulary for a uterine tumor is short: endometrioid or serous or clear cell, grade 1 through 3, FIGO stage I through IV. A single whole-slide image routinely holds more variation than those categories can carry — preserved glandular differentiation in one field, solid sheets in the next, then necrosis, stromal reaction and a dedifferentiated invasive front elsewhere on the same section. SpaTIE treats that coexistence as signal rather than sampling noise. If morphologically distinct regions of one specimen are asynchronous snapshots of tumor states distributed along a progression continuum, they can be ordered, and the ordering recovers a continuous axis that no categorical label encodes.

The framework works in two stages. First it adapts UNI, a general-purpose pathology encoder trained across organs and diseases, to uterine tissue through continued DINOv2-style self-supervised learning — the pretrained weights stay frozen and only low-rank adapter matrices inserted into the query and value projections of each attention block are updated, so organ-specific morphology is learned without discarding general histopathological knowledge. Second, with that adapted encoder used as a frozen feature extractor, SpaTIE imports the trajectory-inference machinery of single-cell biology — principal components, a k-nearest-neighbour graph over embeddings, diffusion pseudotime — and runs it over slide-level and patch-level representations instead of cells. The resulting state values are mapped back onto slide coordinates as spatial tumor-state maps.

SpaTIE was developed by groups at Tsinghua University's Shenzhen International Graduate School, Fuzhou University and the Department of Pathology at Peking University Third Hospital, and posted as a preprint in August 2026. It follows the same organ-specialization strategy the group applied to kidney biopsies in GloPath and RenalPath.

#Key Features

• Parameter-efficient organ adaptation: Only the LoRA matrices on the query and value projections are trained during the 50-epoch continued-pretraining run; the ViT backbone is otherwise untouched, and no downstream task refits the encoder.
• Tumor-focused pretraining corpus: PK3-Uter, an archival uterine cohort of 10,426 H&E whole-slide images, yields roughly 62.3 million 20× patches after quality control, which a tumor-region filter narrows to the 24,318,109 patches actually used.
• Continuous state inference without supervision: Diffusion pseudotime over the embedding manifold produces a tumor-state axis with no temporal sampling, molecular labels or staging information as input.
• Spatially resolved state maps: Patch embeddings are interpolated into the same manifold, so every region of a slide receives a state value and heterogeneity is readable in place rather than averaged away.
• Multi-omics anchoring: Inferred states are correlated against TCGA methylation, copy-number, mutation, RNA-seq and RPPA profiles, and a virtual perturbation score ranks which molecular features carry the association.

#Technical Details

The encoder is a ViT-L/16 with 1024-dimensional embeddings across 24 layers, roughly 303 million parameters, initialized from UNI and adapted with AdamW at a learning rate of 1e-4 on A800 GPUs. Slide-level representations come from attention-based multiple-instance aggregation over CLS-token patch features. All evaluation is external: TCGA-UCEC (566 slides after quality control) and TCGA-UCS (91 slides), with patient-level splits for the supervised probes. On TCGA-UCEC the adapted manifold is the most compact of those compared — 45.833 principal components to reach 90% of the variance, against 48.713 for UNI and 62.272 for an ImageNet-pretrained baseline — while UNI stays more compact on the small TCGA-UCS cohort. Inferred states are stable under perturbation (rank correlation 0.997) and spatially coherent (0.919 and 0.893 across the two cohorts). Clinically, the state axis tracks FIGO stage, tumor grade, invasion depth and fraction genome altered in TCGA-UCEC, and separates overall survival (log-rank p = 0.0038; Cox HR 0.677, 95% CI 0.536–0.855); in TCGA-UCS it correlates with fraction genome altered (ρ = 0.219).

#Applications

The frozen features serve the standard computational-pathology tasks — subtype and grade classification, prediction of TP53, PTEN and ARID1A status, molecular subtyping and survival modelling — from routine H&E alone, with attention maps that localize to tumor-rich regions. The distinctive use is the state axis itself: a continuous coordinate for stratifying patients who share a diagnostic label, and for generating hypotheses about which molecular programs co-vary with morphology before committing to sequencing.

#Impact

SpaTIE is an argument that organ-specific adaptation of a pathology foundation model buys something beyond a benchmark point — a latent space coherent enough to support trajectory inference. The gains over UNI on prediction tasks are modest and task-dependent, and the authors are explicit that the inferred axis is morphology-derived organization rather than a chronological history of any tumor. Sample sizes limit the carcinosarcoma results, the pretraining cohort is single-institution and not public, and the multi-omics associations come from bulk profiles that cannot be assigned to the spatial regions driving each state map. The paper is a preprint; source code and trained weights are announced for a GitHub repository that currently holds no files.