Proscia launches pathology AI model toolkit

Proscia has launched Concentriq Embeddings for experimenting with multiple pathology artificial intelligence (AI) models in parallel.

Embeddings is designed to accelerate the development of prototype biomarker prediction models, according to Proscia. The company has paired Embeddings with its new Proscia AI Toolkit, a suite of open-source resources. Together, the products are designed to enable life sciences organizations in their discovery and development of novel therapies and diagnostics, according to Proscia.

Julianna Ianni, PhD.Julianna Ianni, PhD.

"Concentriq Embeddings efficiently transforms pathology pixels using foundation models to create a digital embedding of images onto an abstract high-dimensional surface," Proscia's vice president of AI research and development Julianna Ianni, PhD, told LabPulse.com. "It turns pixels into a list of numbers that represent aspects (shape, texture, morphology, etc.) of the image," Ianni said. 

These high-dimensional numerical representations are generated from whole-slide images in combination with four-vision and vision-language foundation models -- DINOv2, PLIP, ConvNext, and CTransPath. Researchers can select the best foundation model for their needs, with applications ranging from image classification and segmentation to risk scoring and multimodal data integration.

"We anticipate that researchers will want to have a variety of pathology-specific models to select from while also exploring the power of less-specialized models like DinoV2, which have the advantage of being trained on extremely large datasets," Ianni said.

ConvNext offers users a more traditional model that uses convolutional layers rather than a transformer architecture, which may be of interest to users, Ianni added. Meanwhile, a model like PLIP can be used on downstream models incorporating language as well as vision, opening endless possibilities, she said.

"Research is increasingly demonstrating that there is no one 'best' foundation model for all downstream tasks," Ianni said. "Every task will require experimentation to determine which model is best suited. There have also been recent findings suggesting that even better performance can be achieved by combining multiple foundation model embeddings together."

Proscia noted the Concentriq platform has been further enriched by its real-world data offering.

Page 1 of 19
Next Page