Problem
In cryo-CLEM, someone has to scan 3D cryo-fluorescence volumes by eye and type out region-of-interest coordinates by hand. On large fields that is slow, it depends on who is doing it, and it leaves no machine-readable record. The whole cryo-EM pipeline downstream waits on those coordinates, so the manual step bottlenecks everything.
Approach
I framed ROI selection as a ranking-by-uncertainty problem and built it as four coupled levels: zero-shot detection, semantic scoring, spatial plausibility, then operator curation that feeds decisions back up. The contribution is the coupling, not any single model.
What I built
- Level 1: zero-shot instance segmentation (Cellpose-SAM) on the Z-projection of cryo-fluorescence MRC volumes. No experiment-specific training.
- Level 2: per-ROI semantic scoring with an open-source biomedical VLM (BioMedCLIP), kept behind a narrow interface so the encoder can be swapped.
- Level 3: rule-based spatial plausibility. Boundary truncation, a robust MAD size-outlier test, and isolation relative to the candidate population.
- Level 4: a single-window napari viewer that shows candidates uncertainty-first and records every accept or reject to schema-versioned JSON beside the source volume.
Result
A working, dependency-light open-source tool. It runs on CPU or GPU with no training data, and produces candidate ROIs on a pilot set of cryo-Airyscan mouse-brain volumes for the Klykov Lab. Evaluation is operator-judgment for now; precision and recall come next.