How do you go about profiling and optimizing a performance-critical medical AI system in Rust?
This talk presents a case study of porting a radiology AI system from Python to Rust, a system that had been shown to boost radiologist productivity by up to 40% in a study, but that relies on high-performance ingestion and processing of data to deliver these results. This resulting system leverages both Tokio and Rayon to power a complex pipeline including async, CPU-intensive, and GPU-intensive workloads.
Attendees will learn the practical profiling techniques used to identify bottlenecks in this system, will see the decisions made in order to overcome these bottlenecks, and will get a brief look at those challenges that are still left tackle. The result is a system that processes nearly a million radiology studies a year across 10+ hospitals on one on-prem commodity GPU.
Principal Solutions Architect, Northwestern Medicine
Eric Karl is the technical lead for the ARIES Radiology project at Northwestern Medicine. He has led a push to leverage Rust in Northwestern Medicine's AI efforts.
