HealthTech · AI
AI medical imaging that reads X-rays, CT and MRI scans for a healthcare provider
A deep learning system that flags pneumonia, tumours and fractures on X-ray, CT and MRI scans, so radiologists see the urgent cases first. It runs in the cloud, which lets smaller hospitals in the network get an AI assisted read without a radiologist on site.
70%Faster diagnosis, with patient wait times down from hours to minutes- Client
- A large healthcare provider running hospitals and diagnostic centres across several cities
- Region
- Global
- Sponsor
- Head of radiology and the chief medical information officer
- Duration
- 24 weeks to the first hospital going live, then a phased rollout to remote sites
Engagement model
FDE: Forward Deployed Engineering
Our engineers sat with the radiology and IT teams through data preparation, model training and the hospital rollout, so clinical feedback reached the model every week.
How FDE worksThe team
- Engagement lead1
- Machine learning engineer3
- Backend and integration engineer2
- Data annotation coordinator1
- QA and validation analyst1
The challenge
Where things stood
- Scans were read manually, and the error rate on busy shifts worried the clinical leadership.
- Patients waited hours for a report, which held up treatment decisions in emergency and outpatient care.
- Remote hospitals in the network had little or no access to an experienced radiologist.
- Patient images are sensitive, so any system had to meet HIPAA and GDPR expectations on encryption and anonymisation.
The approach
How we ran it
Weeks 1 to 6
Data and labelling
Pulled historical X-ray, CT and MRI studies from the PACS, stripped patient identifiers and set up a labelling workflow where radiologists confirmed findings for pneumonia, tumours and fractures.
Weeks 5 to 14
Model training
Trained convolutional neural networks per modality and tested them against cases the radiologists had already reported, reviewing every disagreement with the clinical team.
Weeks 13 to 20
Hospital integration
Connected the models to the imaging workflow so a new study is scored on arrival, then placed on a priority worklist with the suspected finding highlighted on the image.
Weeks 20 to 24, then ongoing
Pilot and remote rollout
Ran a supervised pilot at the main hospital, then opened cloud access to remote sites, where local doctors see the AI read while a radiologist reviews remotely.
What it looks like
Representative screens
Drawn to show the kind of system we delivered. Client screens and data stay private.
Study ST-48213, chest X-ray
- Patient
- Anonymised ID P-90321, male, 58
- Referring site
- Ahmedabad main, emergency
- AI finding
- Consolidation, right lower lobe
- Model confidence
- High
- Assigned to
- Dr. Meera Iyer, radiology
- Study received from PACS
- Identifiers removed
- AI read completed
- Radiologist review
- Report sent to ward
What we delivered
Modules and features
- Convolutional neural network models for X-ray, CT and MRI findings
- Anonymisation pipeline that removes patient identifiers before any image leaves the hospital
- Priority worklist that moves suspected critical findings to the top
- Image viewer overlay that marks the region behind each finding
- Cloud deployment so remote hospitals can request an AI assisted read
- Audit log of every model result and the radiologist's final decision
- Monitoring dashboard for model agreement with radiologists over time
What changed
Outcomes
- 70%Faster diagnosis, with patient wait times down from hours to minutes
- Flags suspected pneumonia, tumours and fractures for radiologist review; urgent scans move to the top of the list
- Remote hospitals in the network can now get an AI assisted read on the same day
- Scan data encrypted and anonymised before analysis, with a full audit trail
Figures as reported by the client.
With near-instant and highly accurate analysis, we can now detect diseases early and efficiently.
Solutions
- Hospital managementHospital Management System
- BI and dashboardsBusiness Intelligence and Management Dashboards
Industry and platforms
Looking back
The models were the easier part. Most of the early weeks went into cleaning labels, because historical reports used different words for the same finding and some studies had no confirmed outcome at all. If we did it again, we would put a radiologist on the labelling rota from the first week. Showing the region behind each finding on the image mattered more to adoption than the scores did: radiologists trusted a result they could check in a second, and ignored one they could not.
Start a conversation
Discuss a similar project
Tell us where your team is today and we will walk you through how a project like this would run for you under FDE. The first conversation is free. Delivery work is paid only after you accept it.
A specialist replies within one working day. The first conversation is free.