CASE STUDY
Dice coefficient
NHS-grade clinical AI cutting reporting time by 85%
Dice coefficient
classification accuracy
faster reporting
workflow efficiency gain
A deep-learning CT analysis pipeline for lung nodule risk stratification, reaching a Dice coefficient above 0.85 and 92% classification accuracy, wrapped in role-based clinical workflows. Reporting time fell from around 30 minutes to about 5 — an 85% reduction — with a 40% workflow efficiency gain in pilot data.
Radiologists were spending ~30 minutes per CT report while referral queues grew — clinical attention consumed by admin instead of patients.
We engineered a U-Net segmentation pipeline achieving a Dice coefficient above 0.85, with a hybrid malignancy classifier reaching 92% accuracy for lung nodule risk stratification.
Around the models we built a full role-based clinical platform — pulmonologist, radiologist and admin workflows covering referrals, nodule tracking, MDT discussions and automated decision logging.
Automated 2D/3D visualization and auto-generated PDF/MP4 reports cut reporting time from 30 minutes to ~5 (85% reduction), with pilot data showing a 40% workflow efficiency gain.
What made this hard
A clinician signs the output
Every result reaching a record is somebody's professional responsibility. That is a constraint on the interface and the audit trail before it is a constraint on the model.
Local validation, not published performance
Accuracy moves with scanner, protocol and population. Figures inherited from a paper do not transfer, so validation had to be done on the data the system would actually see.
Reporting is the bottleneck, not detection
The clinical time cost was in producing the report, so automating detection alone would have moved a number nobody was measured on.
Imaging lives behind its own boundary
PACS and DICOM sit on a separate network with their own access path, which shapes where inference can run.
How we built it
U-Net segmentation
Nodule segmentation achieving a Dice coefficient above 0.85 on held-out data from the deploying site.
Hybrid malignancy classifier
92% accuracy for risk stratification, with performance reported by subgroup rather than as a single figure.
Role-based clinical platform
Pulmonologist, radiologist and admin workflows covering referrals, nodule tracking, MDT discussion and automated decision logging.
Automated reporting
2D/3D visualisation with generated PDF and MP4 outputs for clinician review — the step that moved the headline number.
The stack
What should you take from this?
- The model was not the hard part. Clinical sign-off, audit trails and workflow integration were, and they are what made the system deployable rather than interesting.
- Automating the reporting rather than only the detection is why the time saving was 85% instead of marginal. Find the step that actually consumes clinician time.
- Start with the administrative workflow if you want measurable impact this year. Diagnostic work carries a governance burden that a first project rarely survives.
Frequently asked questions
Is this a regulated medical device?
Software influencing a diagnostic or treatment decision may be, and where the line falls decides the entire evidence burden. Establishing that early changes cost and timeline more than any technical choice.
Would these numbers hold at our site?
Not automatically. Performance moves with scanner, protocol and coding practice, which is why local validation is part of the work rather than an optional extra. Any figure not measured on your data is marketing.
How do clinicians interact with it?
Through the systems they already use, with the output presented for review and an override path that is recorded. A tool requiring a separate login does not get used.
Related service: Healthcare AI Development
Build something at this level
Tell us what you are building. We reply within one business day.