RadFlow AI: AI-Powered Radiology Workflow Assistant
Challenge
Radiologists split time across a PACS viewer, a separate AI interface, and dictation reporting, and one-third of reading time went to non-interpretive tasks. Chest CT studies loaded in 8 to 12 seconds at rural satellite sites. The prior commercial AI returned 4.1 false positives per scan, so radiologists dismissed its findings without review. Scan volume grew 22% a year against flat headcount; turnaround ran 15% past contractual SLAs; and the system had to meet HIPAA/HITECH, align to IEC 62304, and stay compatible with an FDA 510(k) pathway
Radiologists split time across a PACS viewer, a separate AI interface, and dictation reporting, and one-third of reading time went to non-interpretive tasks. Chest CT studies loaded in 8 to 12 seconds at rural satellite sites. The prior commercial AI returned 4.1 false positives per scan, so radiologists dismissed its findings without review. Scan volume grew 22% a year against flat headcount; turnaround ran 15% past contractual SLAs; and the system had to meet HIPAA/HITECH, align to IEC 62304, and stay compatible with an FDA 510(k) pathway
Solution
Codebridge built a unified diagnostic workspace with five modules: AI triage worklist, WebGL 2.0 viewer with toggleable overlays, clinical oversight dashboard, immutable audit and explainability logs, and an integration layer. Progressive DICOM streaming over DICOMweb and adaptive bandwidth compression hold initial render under 400ms. A 3D Feature Pyramid Network on a ResNet-50 backbone reads each study in about 47 seconds, with a false-positive reduction network and a longitudinal prior-study comparison module. Grad-CAM saliency maps explain each finding, and one-click adjudication feeds a shadow-mode retraining loop. The stack is React, OHIF, PyTorch, MONAI, NVIDIA Triton, FastAPI, and AWS EKS, delivered under IEC 62304-aligned CI/CD, HIPAA Safe Harbor de-identification, TLS 1.3, and a design history file for a future 510(k). Eight engineers shipped the platform in 24 weeks.
Codebridge built a unified diagnostic workspace with five modules: AI triage worklist, WebGL 2.0 viewer with toggleable overlays, clinical oversight dashboard, immutable audit and explainability logs, and an integration layer. Progressive DICOM streaming over DICOMweb and adaptive bandwidth compression hold initial render under 400ms. A 3D Feature Pyramid Network on a ResNet-50 backbone reads each study in about 47 seconds, with a false-positive reduction network and a longitudinal prior-study comparison module. Grad-CAM saliency maps explain each finding, and one-click adjudication feeds a shadow-mode retraining loop. The stack is React, OHIF, PyTorch, MONAI, NVIDIA Triton, FastAPI, and AWS EKS, delivered under IEC 62304-aligned CI/CD, HIPAA Safe Harbor de-identification, TLS 1.3, and a design history file for a future 510(k). Eight engineers shipped the platform in 24 weeks.
Results
CT reading time fell from 15.2 to 9.4 minutes and P95 turnaround from 6.2 to 3.8 hours. Sensitivity on sub-4mm nodules reached 96% against a 93% threshold, and false positives per scan fell from 4.1 to 0.4. Satellite load times fell to 0.4-0.9 seconds, and radiologist trust rose from 27% to 89%. Uptime is 99.97% over nine months, and a double-blind study on 2,400 scans confirmed the claims. Estimated annual operational impact is about $2.1M.
CT reading time fell from 15.2 to 9.4 minutes and P95 turnaround from 6.2 to 3.8 hours. Sensitivity on sub-4mm nodules reached 96% against a 93% threshold, and false positives per scan fell from 4.1 to 0.4. Satellite load times fell to 0.4-0.9 seconds, and radiologist trust rose from 27% to 89%. Uptime is 99.97% over nine months, and a double-blind study on 2,400 scans confirmed the claims. Estimated annual operational impact is about $2.1M.
