D24AC00338

Current clinical approaches like intraoperative pathology have not solved the persistent problem of incomplete tumor removal.

Several technologies have been tried over the past 20 years, but there is currently no existing technology that can deliver the technical performance needed to fully address this problem and survive in the wild.

In the MAGIC SCAN project, we will introduce key innovations in microscopy, sample automation, cyberinfrastructure, ML model co-design and training on petascale data, practical rapid ML model deployment, and cancer detection and visualization.

We will take a human-centered approach to innovation design and development, involving end users and stakeholders in every aspect of the project to accomplish trustworthy, practical, capable, and cost-conscious product design that optimizes benefits to physicians, payers, and patients.

We will develop the world's fastest high-resolution tissue scanner and use it to:

I. Automatically prepare, handle, and scan the complete surface of removed cancerous organs at 0.5 µm resolution.

II. Detect the presence of any residual cancer cells.

III. Map their location on the specimen surface for surgeon visualization in the operating room within 15 minutes.

Two complementary concepts, each representing fundamental scientific and technical advances in their respective fields, comprise our human-centered design.

The first, MAGIC SCAN machine learning-assisted gigantic image cancer margin scanner, combines extreme field of view, optical sectioning, and super-resolution structured illumination microscopy (OS SR SIM) to obtain very high-speed virtual pathology imaging of cancer tumor margin surfaces at 2x the resolution allowed by diffraction, meeting the PSI requirement and satisfying the TA1 A BAA requirements of area coverage, resolution, and cancer visualization within 10 minutes.

To accomplish the cancer classification goals of an end-to-end solution in less than or equal to 10 minutes, both practically and economically, we propose the companion concept FASTMAP (Fast Accelerated Support for Training Machine Learning Models on Petascale Data).

FASTMAP is a human-centered approach that addresses the iterative nature of the creation-evaluation process of ML cancer classification models by implementing a high-performance computing cyberinfrastructure capable of developing and training new accurate models over petascale data on a scale of days.

We address this challenge using data management and processing techniques developed over almost two decades of high-performance computing research that are deployed in the latest exascale computing environments.

This approach will be complemented by novel strategies for lightning-fast image processing and ML inference at the edge, processing TB of data in minutes within the constraints of rural hospital operating rooms.

Via nine major tasks, we propose to develop an imaging system that can:

I. Provide pathology-quality images over 450 cm² of wet, cauterized, bulky, and irregular fresh tissue at a 250 nm pixel resolution.

II. Deliver automated cancer detection and localization, all within 10 minutes using ML models trained on huge yet challenging datasets.

III. Be practically and relatively inexpensively deployed in complex operating room environments.

If successful, our work would transform cancer surgery as we know it and end the worry of incomplete tumor removal for every American facing a cancer operation.

This would have tremendous health, economic, and societal benefits.

The Administrators Of Tulane Educational Fund

THE FOUR INITIAL FOCUS AREAS ARE(1) HEALTH SCIENCE FUTURES(2) SCALABLE SOLUTIONS(3) PROACTIVE HEALTH(4) RESILIENT SYSTEMS

93.384 - ADVANCED RESEARCH PROJECTS AGENCY for HEALTH (ARPA-H)

Interior Business Center (IBC) [DOI]

Orleans, Louisiana United States

County-Wide

D-AQD-FA-24-023