United StatesHealthcare and life sciences2020 – ongoingJohns Hopkins University BIOS Division
Machine learning review of CT scans in brain hemorrhage research
Johns Hopkins brain injury team cut CT scan review from five hours to 30 seconds per scan with machine learning
Problem
Reviewing brain CT scans from roughly 500 hemorrhage patients took about 2,500 hours across 7 to 10 researchers. The division wanted faster, more accurate image review for its clinical trials.
Approach
The Johns Hopkins Brain Injury Outcomes Division trained machine learning models on its library of brain hemorrhage images, working with Google Cloud partner Quantiphi. Data came through an NIH cloud program. The team used 3D U-Net models trained on Google Cloud and ran the inference pipeline on Kubernetes.
Results
Takeaway
The division said the machine reviews were also more accurate than human review of the same images. It described a 30-second result at the point of care as something the approach could allow, not something already in use.
Categories
- Size
- Large enterprise
- Industry
- Healthcare and life sciences
- Business function
- IT and R&D
- Innovation type
- Process innovation
- AI technology
- Computer vision
- Adoption
- Partner co-development
- Evidence
- Vendor case study
Sources
Source summaries
Google Cloud's customer story describes how the Johns Hopkins BIOS Division and Quantiphi used machine learning to review brain CT scans. It reports review time falling from five hours to 30 seconds per scan, and from 2,500 hours to 4.5 hours for about 500 patients. The page carries no date; the Internet Archive first saved it on October 4, 2020.
Full details are in the original sources.