BlueSpotter — LC-optimized Cellpose-SAM
Fine-tuned Cellpose-SAM weights for instance segmentation of locus coeruleus (LC) neurons in neuromelanin / histology microscopy slices. Part of the BlueSpotter project (priggelab.com): a standardized, hardware-independent pipeline for quantitative LC analysis.
What this is
- Base model: Cellpose-SAM (
cpsam). - Task: instance segmentation — each detected neuron gets a unique integer label.
- Trained on: paired image/mask crops from neuromelanin slice cohorts (see the BlueSpotter repo's data convention and manifest).
Usage
from huggingface_hub import hf_hub_download
from cellpose import models
# Download the fine-tuned weights from this repo
weights = hf_hub_download(repo_id="TeamPrigge/bluespotter-lc", filename="bluespotter_lc")
# Load exactly like the base Cellpose-SAM model, but with these weights
model = models.CellposeModel(gpu=True, pretrained_model=weights)
# Segment — masks is a label image; masks.max() == number of neurons found
masks, flows, styles = model.eval(image)
n_neurons = int(masks.max())
The inference call matches BlueSpotter's training-time QC (src/bluespotter/viz.py),
so masks here are directly comparable to the validation overlays produced during training.
Intended use & limitations
- Intended: research-grade segmentation of LC neurons in similar imaging modalities to the training data.
- Out of scope: clinical/diagnostic use; cell types or modalities far from the training distribution. Always visually QC predictions before quantifying.
Training
Fine-tuned via the BlueSpotter Colab workflow. Hyperparameters and per-epoch losses
for each run are tracked in MLflow; see params.yaml and docs/SETUP.md in the
code repository.
Citation
If you use these weights, please cite Cellpose-SAM and credit the BlueSpotter project (priggelab.com).
License note:
cc-by-nc-4.0is a placeholder default for non-commercial research sharing — confirm or change it before publishing.