← All workBachelor's thesis · 2026
Finding brain aneurysms in MRA
A two-stage 3D pipeline that finds cerebral aneurysms a few millimetres across inside full TOF-MRA volumes, validated on 974 external series.
- 91.64%
- specificity
- 974
- external series validated
- 0.656
- best Dice inside the crop
Case 01 · de-identified test series
A full TOF-MRA study is 512 × 512 × 112 voxels. An aneurysm is a few dozen of them. Train a single network on that and it learns that “no aneurysm” is almost always the right answer.
So the thesis splits the problem in two. A first nnU-Net (3d_fullres, residual encoder) runs on the whole volume and only has to propose candidates. Each connected candidate becomes a centred 128³ crop, and a second network, trained only on those crops, does the actual segmentation.
Here is what that looks like on the test case: the real slices around the detection, with the second network’s segmentation on top.
On top of that, an optional vessel mask from VesselVerse rejects components that are anatomically impossible: an “aneurysm” floating outside any vessel is a false positive by definition.
What I’d point out
- Each operating point is reported on its own. It’s tempting to quote the sensitivity of one configuration next to the specificity of another. I kept them separate.
- External validation on 974 series from a different source than the training data.
- A local review tool on top of the precomputed, de-identified outputs, so a clinician can check a result in context without a GPU.