Beyond the Scan
| Context | AP Research |
|---|---|
| Role | Individual researcher |
| Dates | August 2025 – May 2026 |
Medical imaging models are usually benchmarked on datasets far larger than a hospital studying a rare tumor will ever assemble. I asked a narrower question: as training data shrinks toward a realistic low-data scenario, which ensemble architectures degrade gracefully, and where is the floor?
I benchmarked five ensemble architectures on rare brain tumor MRI classification, progressively shrinking the training set and measuring how each held up. DenseNet169-LIME-TumorNet proved the most robust under scarcity, and the curves put a practical data-sufficiency threshold at roughly 28 training images. Below that, performance fell away sharply enough that more data matters more than a better architecture.
The LIME component is doing real work here rather than sitting in the name. In a clinical setting a classification that cannot be interrogated is close to useless, because the person acting on it has to decide whether to trust it on this particular scan. Carrying local interpretability through the ensemble meant the robustness result and the explainability result could be read together instead of traded against each other.
The finding I care about is the threshold, not the leaderboard. Knowing roughly where a model stops being data-limited tells a small research group whether their next move is collecting more scans or changing their approach, which is the decision they actually face.