Source research

A multimodal whole-slide foundation model for pathology

Tong Ding, Sophia J. Wagner, Andrew H. Song, Richard J. Chen, Ming Y. Lu, Andrew Zhang et al. · 2025 · 151 citationsRead the paper

The field of computational pathology has been transformed with recent advances in foundation models that encode histopathology region-of-interests (ROIs) into versatile and transferable feature representations via self-supervised learning. However, translating these advancements to address complex clinical challenges at the patient and slide level remains constrained by limited clinical data in disease-specific cohorts, especially for rare clinical conditions. We propose Transformer-based pathology Image and Text Alignment Network (TITAN), a multimodal whole-slide foundation model pretrained using 335,645 whole-slide images via visual self-supervised learning and vision-language alignment with corresponding pathology reports and 423,122 synthetic captions generated from a multimodal generative AI copilot for pathology. Without any fine-tuning or requiring clinical labels, TITAN can extract general-purpose slide representations and generate pathology reports that generalize to resource-limited clinical scenarios such as rare disease retrieval and cancer prognosis. We evaluate TITAN on diverse clinical tasks and find that it outperforms both ROI and slide foundation models across machine learning settings, including linear probing, few-shot and zero-shot classification, rare cancer retrieval, cross-modal retrieval and pathology report generation. Pretrained using 335,645 whole-slide images, a foundation model is developed to provide representations for slide- and patient-level tasks. It is capable of performing clinical tasks and generating reports even in data-scarce scenarios, such as rare cancer diagnosis and survival prediction, without requiring further fine-tuning.

What the full text shows

Working prototype

TITAN outperforms other slide encoders (PRISM, GigaPath, CHIEF) across morphological classification, grading, molecular classification, and survival prediction tasks. (Linear probing on frozen slide embeddings)
TITAN demonstrates superior parameter efficiency compared to PRISM and GigaPath. — 48.5M parameters vs 99.0M (PRISM) and 86.3M (GigaPath) (Model architecture comparison)
Increasing pretraining data scale improves performance on subtyping tasks (TCGA-UT-8K, TCGA-OT, OT108, EBRAINS). — Average increase of 3.65%, 3.21%, and 1.21% when scaling from 12.5%, 25%, and 50% of the full dataset (TITAN V pretraining)
Linear probing of TITAN embeddings outperforms supervised attention-based MIL (ABMIL). (Downstream clinical tasks)
Fine-tuning TITAN from pretrained weights improves performance over linear probing, whereas fine-tuning from random weights decreases it. — -3.63% average performance for random initialization (Task-specific fine-tuning)
TITAN significantly outperforms other encoders in low-data regimes. — P < 0.0001 (Few-shot learning settings)

Setup: Pretrained on Mass-340K dataset (335,645 WSIs, 182,862 reports, and 423,122 synthetic captions) across 20 organs.
How: A three-stage pretraining approach: (1) vision-only iBOT SSL on ROI crops, (2) vision-language alignment of ROIs with synthetic captions, and (3) alignment of WSIs with pathology reports. It uses a ViT architecture with 2D ALiBi positional encoding to handle long-context extrapolation of pre-extracted CONCHv1.5 patch features.
Limitations: Pretraining on 8k x 8k region crops and extrapolating with ALiBi may not capture full contextual information of the entire WSI.

Summarized by Gemma 4 from the open-access full text · 10 Oct 2026. Read it