Putting a deep-learning model in clinicians’ hands
An end-to-end serving application that let clinicians run inference, flag edge cases and re-annotate data — closing the loop between model output and model improvement.
- Client
- IIT Ropar, with PGIMER Chandigarh and Dr. Sukrit Gupta (NTU Singapore)
- Role
- Full-Stack Research Engineer (Intern)
- Timeline
- May 2024 – Dec 2024
- Status
- Deployed to clinical workflow
The gap between a model and a tool
A segmentation model that only its authors can run is a research artefact, not a clinical tool. The gap is not accuracy — it is that a clinician has no way to run it on a real case, see where it went wrong, and say so.
The project set out to close that gap: give clinicians a reliable interface to the model, and give the model a route back from clinical judgement into better training data.
What I built
An end-to-end deep-learning serving application with a GUI clinicians could actually use. Run inference on a case, inspect the output, flag edge cases where the model failed, and re-annotate the data directly.
That last step is the one that matters. Every correction a clinician makes becomes training signal, so the system improves through use rather than through another round of data collection.
Solo MVP, then a team
I built and tested the initial MVP solo to prove the workflow held together, then led and mentored a remote team of student developers to scale it into something that fits a real clinical workflow.
Doing both halves taught me where the seam is: the decisions you make alone in week two are exactly the ones that determine whether five people can work on it in month four.
Working with domain experts
I partnered with Dr. Sukrit Gupta (NTU Singapore) and clinical partners at PGIMER Chandigarh to meet stringent medical standards. The stack — MONAI, nnUNet, UNet, CVAT and 3D Slicer — was chosen to sit inside tooling clinicians and researchers already trusted rather than asking them to adopt something new.
Most of the engineering difficulty in domain-expert projects is not technical. It is learning enough of someone else’s field to ask the right question, and being specific enough that they can correct you.
Stack
Need something like this built?
I take on freelance briefs and part-time engineering work from Melbourne, Australia.