Matt Lee, Director of AI and Medical Imaging at Sonrai, delivered a presentation on unlocking prognostic biomarkers within digital pathology, focusing on the integration of foundation models. Lee began by introducing Sonrai, a Belfast-based company dedicated to advancing precision medicine through connected data and collaborative science. The Sonrai Discovery platform, he explained, provides a trusted research environment for multiomic, clinical, and imaging data, ensuring auditability and reproducibility.
Lee outlined the current landscape of computational pathology, highlighting its impact on biomarker discovery, tumour microenvironment analysis, and clinical decision support. He emphasised the challenges faced in the field, such as variability in sample preparation, staining, and scanning, which can hinder reproducibility. Regulatory approval and external validation remain significant hurdles, as do issues of interpretability, bias, and generalisability – particularly when narrow training datasets are used. Lee also discussed the complexities of integrating pathology data with other modalities, such as genomics and clinical data.
The core of the presentation centred on the use of foundation models in digital pathology. Lee described how these models generate feature vectors or image embeddings from pathology images, enabling clustering and annotation at both tile and whole-slide levels. He illustrated how clustering can facilitate the labelling of tissue types and morphologies, while whole-slide embeddings provide a condensed, fixed-size representation of entire images, regardless of their original scale. This approach, he argued, addresses the challenges of data scale and variability, and is highly amenable to integration with multiomic and clinical data.
Lee presented evidence that pathology-specific foundation models outperform generic models, offering greater generalisability and requiring less data to achieve state-of-the-art results. He demonstrated the utility of whole-slide embeddings in survival modelling using public datasets, showing clear risk stratification and statistically significant results. Lee concluded that whole-slide embeddings are a powerful tool for the future of multimodal precision medicine, noting recent regulatory milestones such as the FDA’s authorisation of multimodal AI for prostate cancer. He encouraged further exploration and collaboration in this rapidly evolving field.
