Data Day Chest x-ray TubeLocate
Transcript of Data Day Chest x-ray TubeLocate
Data DayChest x-ray TubeLocateMay 27th, 2021
Dheeraj Boda (OHIA –Solution Architecture and Engineering)
Introduction and the Ask
• Design and implement a solution to Productionalize the Real time image processing Application and Models using scalable, efficient and robust processes and infrastructure
• Have DR and HA enabled secure environments for image processing• Have automated roundtrip processing time less than 4 min• Store logging and intermediate resulting information, make available for
further investigation, enhancements
Enabled Azure Services • Azure Storage• Azure Private Endpoints• Azure Key Vault• Azure Event Grid• Azure Functions• Azure Batch Service• Azure Kubernetes Service• Azure SQL DB• Azure Application Insights• Azure Machine Learning service
Future Enhancements with XOpsXOps is an umbrella term used to define and integrate various aspects of automating machine learning lifecycle that includes various IT disciplines and responsibilities which could include Development, deployment, security, networking etc. with IT operations to be secure, reliable and efficiently maintained.
DevSecOps
1.Data Pipelines2.Environment Pipelines3.Model Training Pipelines4.Model Deployment
Pipelines5.Azure ML Service (MLflow)6.Azure Databricks (MLflow)
MLOps
UCLA Data Week
Real-Time Intelligence System for Detection of Tubes and Lines in Chest X-Rays
Matthew Brown, PhDCenter for Computer Vision and Imaging Biomarkers
UCLA PACS Datcard RouterCVIBSENN
Microsoft Azure
CareConnect
2) CXR image 3) CXR image
5) Original CXR + Enhanced CXR with AI annotations and alerts
4) DICOM image with AI overlay
UCLA Clinical DeploymentX-ray Unit
1) CXR image
4) DICOM image with AI overlay
IRB for R&D
Train and test AI
FDA 510(k)? No
QI Project? Yes
Integration
Legal Review
Risk Assessment
Rollout – Phase 1
Rollout – Phase 2
Pilot users• Selected physicians that place orders• Selected ACI radiologists• New order number
Weekly meeting during pilot phase
Data set: 1,900 annotated chest x-rays• 1396 training, 514 test
Detection accuracy: 96% ET, 94% NG
Mean ET tip error: 7mm• 86% < 10 mm