We built edge AI for a high-volume hospital that analyses CT and MRI scans on site, locates likely tumours and moves urgent cases up the worklist. Throughput rose 4.5x, and each on-site analysis takes 40ms.
Client: MedTech InnovationsIndustry: HealthcareService: AI & Automation
Overview
Client
MedTech Innovations
High-volume multi-specialty hospital
Engagement snapshot
The challenge
Radiologists worked from a shared PACS (picture archiving and communication system) and reviewed every slice of high-resolution CT and MRI scans by hand. At peak times the backlog reached 18 hours, which delayed urgent treatment.
Cloud AI was not an option because of data residency rules and HIPAA risk, and radiologists were losing a large part of each shift to sorting the worklist.
Urgent findings were buried in long worklists, which delayed post-procedure reviews.
Cloud AI options raised legal questions about data residency and HIPAA risk.
Radiologists spent 40% of each shift on manual prioritisation instead of diagnosis.
The solution
A local streaming gateway picks up DICOM studies straight from the scanners, triggered by HL7 events.
Vision Transformer models run on local edge hardware with quantised weights, keeping 98%+ accuracy with responses under 40ms.
The PACS worklist is re-ordered automatically, and scans with likely tumours are flagged for a radiologist to review straight away.
Key features
Quantised vision transformer models run on local hardware and analyse each scan as it completes to locate likely tumours.
Business impact
98.2% localisation accuracy
Flagged scans move to the top of the worklist, and a radiologist confirms every AI finding.
Business impact
Severe-case wait cut from 18 hours to under 2
Custom quantisation cut the compute footprint by 70% while keeping sub-millimetre localisation accuracy.
Business impact
40ms local inference
A local scrubber pseudonymises data before it enters internal diagnostic logs, meeting HIPAA Title II requirements.
Business impact
100% data control and residency
The system doesn't stop at detection. It works with the radiology information system (RIS) to re-prioritise the radiologists' queue.
Business impact
4.5x prioritisation throughput
Before and after
Before
Serial human review
Radiologists read images in the order they arrived, however severe the case.
After
Parallel edge inference
AI screens every slice as the scan completes and moves high-risk cases forward.
Before
Restricted by policy
Cloud processing was ruled out by patient data transit risks and data residency laws.
After
No data in transit
Inference and storage stay inside the hospital's local network, locked to HIPAA requirements.
Before
Minutes to hours
There was always a wait between a scan finishing and a radiologist's first look.
After
Milliseconds
Tumour bounding boxes are ready before the patient leaves the scanner.
Results
40ms Inference Time
At 40ms per inference, the quantised ViT models matched senior radiologists' consistency on initial localisation work.
Before: VariedAfter: 98.2% Accuracy
4.5x Prioritisation Lift
AI prioritisation cut the wait for likely severe cases by over 80%.
Before: 18h DelayAfter: < 2h Delay
Zero Transit Response Time
Running on local hardware with no data in transit, diagnostics stayed available through hospital-wide internet outages.
Before: 95.0%After: 99.99%
Testimonial
“Severe cases used to wait 18 hours for a read. Now it is under 2. Flagged scans move to the top of the worklist with a heatmap, localisation accuracy is 98.2%, and a radiologist still confirms every finding.”
Eleanor VaneChief of Radiology, MedTech InnovationsKey results
Trust and control
Each AI finding comes with an activation heatmap, so radiologists can check why the model flagged it.
COMPLIANCE READYNothing is sent off site, so all patient health information stays inside the hospital firewall.
HIPAA-ALIGNEDA radiologist must verify each AI result before it is committed to the clinical worklist.
HUMAN IN THE LOOPThe hospital owns 100% of the edge AI runtime and the tuned vision models.
100% OWNEDFaster delivery
These tested components saved about 10 weeks of setup work, which we put into tuning the vision models and the worklist logic.
Our tested data module handled pseudonymisation and local storage to HIPAA requirements from the start.
A reusable DICOM streamer gave us the gateway that pulls each study from the scanner as it completes.
An existing monitoring platform tracked inference times and model accuracy on the edge hardware.
Pre-built reliability controls kept local inference running through hospital-wide internet outages.
A local gateway collects each DICOM study as the scan finishes, and quantised Vision Transformer models analyse every slice on hospital hardware in about 40ms. Likely tumours are marked with bounding boxes and an activation heatmap. Urgent cases then move up the PACS worklist for a radiologist to review.
Prioritisation throughput rose 4.5x, and the delay for likely severe cases fell from 18 hours to under 2. Before, radiologists read scans in arrival order and spent 40% of each shift sorting the worklist. Now the system re-orders the queue so urgent scans are seen first.
Cloud processing was ruled out by data residency laws and the HIPAA risk of sending patient data off site. Running on local edge hardware keeps all patient health information inside the hospital firewall. It also kept diagnostics running during hospital-wide internet outages, with availability rising from 95.0% to 99.99%.
If data residency rules stop you using cloud AI, we can help you run imaging models on your own hardware, next to the scanners.