How Rad AI is Revolutionizing Radiology with Ray
summary
Rad AI automates radiology impressions with Radi Omni, generating ~100,000 impressions/day for ~4,000 daily radiologists and claiming up to 90% fewer words and about a 3s median generation time; edits feed back as training signals. Training infrastructure uses Ray Clusters, Ray Train, Kuberay, Pulumi and GKE plus multi-node A100/H100 setups (production trains on 4–8 A100s), cutting epoch time ~3 hours and reaching comparable accuracy ~20% sooner. Internal orchestration is via rad train and a Kubernetes job queue for cloud‑agnostic, region‑specific training.
AI-written summary · about
chapters (9)
- 0:00Introduction to AI in Rad
- 1:43Automating impressions
- 4:28Demo of Rad AI Omni
- 6:42Full report generation
- 9:57ML training infrastructure
- 11:51Evolution of infrastructure
- 19:14Rad train and workflow
- 23:04Multi-cluster scaling
- 25:38Q&A session
related talks
35mOperating a Production-Scale Physical AI Data Platform with Ray
28mScaling AI & Machine Learning Workloads With Ray on AWS, Kubernetes, & BERT
22mRay, a Unified Distributed Framework for the Modern AI Stack
38mTokio, Rayon, and the GPU: Profiling a Radiology AI Pipeline in Rust
25mAI In Production Key Challenges and Use Cases
46mPath of Data & AI Challenges
29mBuilding Self-Serve, Multimodal Medical Image Pipelines with Ray
24mBuilding Complex Data Analytics Pipelines with Ray