Context and role
| Context | Research collaboration · Babeș-Bolyai University and UMF Iuliu Hațieganu · 2024 to 2025 |
|---|---|
| Role | Built the detection and classification pipeline, working with Prof. Laura Diosan and clinical collaborators. |
| Category | Research |
The problem
Biomarker prediction from lung CT is a 3D problem. Nodules are small, irregular and surrounded by anatomy that produces false positives.
EGFR and KRAS mutations matter for treatment decisions, but the signal is spread across imaging features that are hard to hand-engineer.
A pipeline that finds every nodule is not useful if it buries the radiologist in false positives. Precision and recall have to be balanced against clinical workflow.
Approach
- Used NVIDIA VISTA-3D for 3D segmentation of lung nodules.
- Built a custom ResNet-18 classifier for the EGFR/KRAS prediction task on the segmented volumes.
- Worked with the LIDC-IDRI dataset in a cross-institution collaboration with clinicians.
- Measured recall, false-positive reduction and AUC rather than a single accuracy number.
What I did not claim
The boundary of a project is part of its evidence. These are the claims this work does not support.
- This is a research collaboration, not a deployed clinical diagnostic product.
- The pipeline was validated in a research setting and later confirmed through clinical validation; it is not a medical device.
- The work was collaborative. The clinical and academic partners contributed domain and validation expertise.
- No claim of use on live patient data in a hospital deployment.
Results
- 92% recall on the nodule detection task.
- An 80% reduction in false positives.
- An AUC of 0.87 for the classification task.
- Clinical validation later confirmed the results.
- The project is evidence that rigorous measurement, not just model choice, is what makes a medical AI result defensible.
Stack
What this means for you
The useful part of this project is not the artifact. It is the method: define the measurement first, keep the scope honest, and hand over something your team can run. That method is what transfers to a new problem.
If your situation looks similar, the related services below are the place to start. If it does not, send a short description and I will tell you whether I am the right person.
Page last updated: October 2026