Demo Hospital · Emergency radiology
--:--Night shift
Ready

CT scanner

Every head CT reaches CTFast automatically. In this demo, you choose the scan that arrives.
DICOM slices or a .zip of them

Radiology worklist 8 scans

# Case Exam Waiting CTFast Status

The worklist the radiologists already use, read from the top. Only A17 is a real scan. CTFast can move a scan up but never down, and a radiologist still reads every scan.

About CTFast

A bleed in the brain needs treatment within the hour, but at night a head CT can sit in the reading queue behind scans that arrived earlier. CTFast makes sure the right radiologist sees a likely bleed first, and keeps escalating until someone does. A radiologist still reads every scan and makes every diagnosis.

How a scan flows through the hospital

  1. The scan arrivesThe CT scanner sends every finished head CT to the image archive, and CTFast receives a copy automatically. Nobody presses a button.
  2. The detector scores itIn a few seconds, with a heatmap of the regions that raised the score.
  3. The worklist changesA suspected bleed goes to the top, a possible bleed moves up, every other scan stays where it was.
  4. The right person is alertedThe on-call radiologist's phone shows the key slice and a case brief with sources. No answer in 5 minutes: the backup. In 10: the consultant.
  5. The radiologist reads itOne tap confirms or rejects the flag. A confirmed bleed goes straight to the emergency department and neurosurgery.
  6. Everything is loggedResponse times and false alarms are reviewed, and the thresholds tuned.

On the live board, the CT scanner panel stands in for step 1: you choose which scan arrives. The worklist, the case view and the phones show steps 2 to 6.

What is real in this demo

  • Real: the detector, an EfficientNet-B0 network trained on about 16,000 head CTs from the RSNA brain bleed dataset, running live on real head CTs from CQ500. The heatmap (Grad-CAM), the scores, the two thresholds, the playbook rules and the worklist order are all computed here.
  • Simulated: the hospital, its staff, the other scans in the worklist, the patient records, the phone messages and the timings. No real message is sent.
  • Sample: the case brief is assembled from the simulated record and a small sample library of protocols. A hospital would connect its own approved documents.

Hospital systems would use the same API as this page: API documentation.

Limitations

  • Tested on public data only. On CQ500, alerts catch 91% of bleeds, or 99% with the possible-bleed tier, and clear 91% of normal scans. SGH reports 98% clearance for iScreen on its own scans. A hospital would run CTFast silently on its own scans and tune both thresholds before it alerts anyone.
  • Bleeds inside the brain tissue are missed most often: alerts catch 90% of them on CQ500.
  • PNG or JPEG uploads carry only the brain window, and scores on them run high. Upload DICOM for a reliable score.
  • This is a student project. It is not a medical device and must not be used for patient care.

Data

  • Training: RSNA Intracranial Hemorrhage Detection, Flanders et al. (2020).
  • Testing and the example scans: CQ500, Chilamkurthy et al. (2018), CC BY-NC-SA 4.0.
  • Extra test: Brain CT Hemorrhage, Helwan et al. (2018), Kaggle, CC BY 3.0.