PathAI Announces PLUTO: a State-of-the-Art Foundation Model to Enhance AI-powered Pathology Tools
Foundation models were developed on numerous diverse data and will enable drug discovery and anatomic pathology laboratory use cases
PathAI, a global leader in AI-powered pathology, launched Pathology Universal Transformer (PLUTO), a pathology-centric foundation model, to further differentiate their portfolio of products geared towards biopharma and pathology laboratory customers.
In machine learning, a foundation model (FM) is a model trained using self-supervised learning on a large scale of unlabeled data to mathematically capture salient information from inputs to the model. Once trained, the FM can be adapted to enable specific tasks in new contexts, especially including previously unforeseen contexts and tasks.
The process of adapting a high-quality FM yields better downstream pathology tools, with lower development time and cost, than building bespoke tools from scratch. For example, a pathology FM may be adapted variously to perform cell and nucleus segmentation at microscopic scale on immunohistochemistry (IHC) slides, or to perform histological subtyping at macro-scale for H&E slides. Developing high-quality FMs is particularly important and complex in pathology where there are many different kinds of tasks at different scales of the whole slide image (WSI) data including cell, tissue, and (whole) slide levels.
PathAI’s PLUTO was developed to enable this wide diversity of pathology tasks, and was trained using hundreds of millions of unlabeled image patches from about 160,000 WSIs across 30+ disease areas, numerous indications, stains, scanner types and magnification, and over 50 sources of data. To learn high-quality representations of the unlabeled training data, PathAI researchers designed a pathology-focused self-supervised training process using multi-scale, flexible deep vision transformer architectures with novel training objectives. Researchers then evaluated the quality of the resulting FM by adapting it to a wide diversity of pathology-specific tasks, and found that PathAI’s FM outperforms current state-of-the-art models on (whole) slide, cellular and subcellular tasks. Additionally, PathAI’s FM enables building AI-powered pathology tools at multiple scales and resolutions, and is significantly more compact than previously published models, driving massive reductions in training and inference costs, and enabling highly scalable data generation and product development.
PLUTO and its adaptations are set to power PathAI’s best-in-class suite of AI products for diagnostic pathology labs and biopharma use. PLUTO will be adapted to develop the next generation of PathAI’s best-in-class products. PLUTO’s embeddings will enable prediction of underlying molecular alterations driving cancer and disease phenotypes–including those previously hidden to AI–as well as the discovery of new biomarkers and therapeutic targets.
“PLUTO is meticulously designed to fuel research efforts for better diagnostic and prognostic tools across multiple disease areas. This technology will enable users to compress their pathology workflows using an exceptionally diverse, multi-resolution foundation model at significantly greater speed and a much broader scale than before,” said Harsha Pokalla, Head of Machine Learning at PathAI. “Our aim is to deploy this single, robust model to address a diverse range of applications and use cases within pathology and to continue developing new high performance AI tools that serve our mission to improve patient outcomes.”
