Healthcare & Bio AI
Whole-slide pathology model trained on 335,645 images outperformed region-of-interest and slide foundation models on clinical tasks without fine-tuning
In a 2025 paper in Nature Medicine, researchers describe a model called TITAN that learns from whole-slide images and the language of pathology reports. A whole-slide image is a digital scan of an entire microscope slide. The team trained TITAN on 335,645 such images and aligned it with corresponding pathology report text and 423,122 synthetic captions made by a multimodal generative AI copilot for pathology. The model then performed clinical tasks such as rare cancer retrieval and cancer prognosis without extra fine-tuning or clinical labels. In the study's evaluations, TITAN outperformed region-of-interest and whole-slide foundation models across several machine-learning settings.
What kind of model is TITAN?
TITAN is a foundation model, a large model that learns general-purpose features from many examples and can be applied to several tasks. It uses a transformer design. The training combined visual self-supervised learning, where the model finds patterns in images without hand-labeled examples, with alignment between images and text. That alignment helps the model connect what it sees in a slide image with how pathologists describe it in a report.
How did the researchers test it?
The researchers tested TITAN without fine-tuning it or giving it clinical labels. The test settings included linear probing, a simple way to read features from a model. They also included few-shot and zero-shot classification, where the model sees few or no training examples for the target class. The settings also included rare cancer retrieval, cross-modal retrieval and pathology report writing. Across these settings, TITAN outperformed both region-of-interest and slide foundation models.
Common questions
Does TITAN need extra training for a new disease?
No, according to the study. The model performed tasks in data-scarce scenarios without fine-tuning and without clinical labels.
What are the synthetic captions?
Synthetic captions are text descriptions produced by a multimodal generative AI copilot for pathology, a tool that works with both images and text. The study used 423,122 of these captions during training, along with corresponding pathology report text.
Did the study compare TITAN with other models?
Yes. The study reports that TITAN outperformed both region-of-interest and slide foundation models across the machine-learning settings in the evaluation.
Limits and what's next
These results come from one model and the task evaluations the abstract describes. The training data were 335,645 whole-slide images and 423,122 synthetic captions. The abstract does not say that the model was tested in live clinical care, and it reports no independent replication of the results. It also does not give the size or setting of the patient data behind the clinical task results. The abstract does not state a next step.
Source: Tong Ding et al., “A multimodal whole-slide foundation model for pathology,” Nature Medicine, 2025.