CellposeCellCounter segmentation models

Weights for CellposeCellCounter, which counts cells and scores viability from a phone photograph of a hemocytometer.

Files

file	used by the app	description
hemocytometer_retrained_20260825.npy	yes	Cellpose-SAM fine-tuned for hemocytometer counting. This is the current model.
generalmodel.npy	yes	General-purpose model used for the confluency tab only. Not retrained, not evaluated on this imaging setup.
hemocytometer_v1_baseline.npy	no	The previous hemocytometer model, kept so the comparison below can be reproduced.

Training — hemocytometer_retrained_20260825.npy

Fine-tuned from the built-in cpsam (Cellpose 4.1.1) on 13 images from a phone-adaptor hemocytometer setup, 1,069 hand-corrected cell masks. Images were cropped to a single 4×4 counting block and downscaled to the app's working size (max side 1024) before annotation, so training and inference see identical geometry.

python -m cellpose --train --dir train --test_dir test --mask_filter _seg.npy \
    --use_gpu --learning_rate 0.00001 --weight_decay 0.1 --n_epochs 60 \
    --train_batch_size 1

Annotation was human-in-the-loop: the previous model's output was corrected rather than drawn from scratch. Across the first eight images that took 336 deletions and 79 additions — the base error was over-segmentation of debris, not missed cells.

Evaluation

Six held-out images, 556 hand-annotated cells, one-to-one IoU matching at 0.5. These images were used for neither training nor any parameter choice.

model	count error	precision	recall	F1
retrained	−1.1%	0.940	0.926	0.933
v1 baseline	+40.6%	0.656	0.932	0.768
Cellpose-SAM, off the shelf	−63.7%	0.741	0.238	0.319

False positives fell from 260 to 36 with recall unchanged. Off-the-shelf Cellpose-SAM is not usable on these images — on one field it returned a single object out of 74.

Intended use and limitations

• Brightfield hemocytometer images from a phone adaptor, cropped to one counting block. Performance on whole uncropped frames is much worse: cells fall to ~7 px and the populations stop being separable.
• generalmodel.npy has not been retrained or evaluated here. Confluency results should be treated as indicative.
• Viability is not produced by these models. The app determines it from a local-contrast threshold, described in the Space README.
• Trained on one cell line, one adaptor and two phones. Generalisation beyond that is untested.

Provenance

generalmodel.npy and hemocytometer_v1_baseline.npy are copies of generalmodel.npy and hemocytometermodel.npy from myang4218/cellposemodel (Apache-2.0), mirrored here so the published pipeline does not depend on an external personal account.