AI and the Future of Digital Pathology
Pathologists study tissue samples under a microscope to diagnose diseases and monitor their progression. Advances in digital pathology now allow entire tissue slides to be scanned into very large whole-slide images (WSIs). These high-resolution images contain detailed morphological information, but their size and complexity make the analysis with traditional AI methods challenging.
How TITAN Learns from Images and Text
TITAN, short for Transformer-based pathology Image and Text Alignment Network, is a multimodal foundation model that learns from both visual and textual data. It was trained on over 335k digital slides from 20 organs using self-supervised learning, which allows AI to identify patterns without manual labeling. The model also incorporated synthetic text descriptions generated by an AI assistant, enabling it to learn from a larger and more diverse dataset and improving its ability to generalize across different research tasks.
By linking image features with corresponding textual information, TITAN produces general-purpose digital representations of tissue slides, which can be applied to many research applications without additional fine-tuning. It can identify disease patterns, assist in recognizing rare diseases, predict cancer outcomes, and generate text summaries similar to pathology reports.
In tests across multiple research tasks, TITAN outperformed existing AI models that focus on smaller image regions. It was particularly effective at analyzing rare diseases and connecting visual findings with textual reports, demonstrating the potential of multimodal AI to support scientists in interpreting complex biological data.
Towards AI-Assisted Disease Understanding
By integrating visual and textual information, TITAN shows how large-scale AI systems can capture both the structure and context of disease. In the future, such models could help pathologists and medical researchers analyze tissue samples more efficiently and contribute to the development of new approaches for diagnosis and treatment.
The TITAN model and its source code are publicly available to the research community at https://github.com/mahmoodlab/TITAN and https://huggingface.co/MahmoodLab/TITAN.
Original Publication
Ding et al., 2025: A Multimodal Whole Slide Foundation Model for Pathology. Nature Medicine. DOI: 10.1038/s41591-025-03982-3
