Modelpublic

einarolafsson/toxoplasma-from-cellmask-cpsam

sourceHugging Facemitupdated 8d agoView on Hugging Face

0likes

Model Card

Toxoplasma from Cell Mask (cross-channel)

Segments Toxoplasma gondii parasitophorous vacuoles from the host cell mask channel alone — no parasite-specific stain required. A cross-channel model: it is given the host cell image and predicts where the parasites are.

—Architecture: Cellpose-SAM (cpsam_v2)
—Model Zoo key: toxoplasma_from_cellmask_v1
—Checkpoint: toxoplasma_from_cellmask_pv
—Trained by: einarolafsson

Use it in spaCR

This model is distributed through the spaCR Model Zoo. spaCR is an open-source package for spatial phenotype analysis of CRISPR screens and microscopy images.

bash

pip install spacr

Model Zoo (GUI)

Launch the GUI and open the Model Zoo:

bash

spacr

Find Toxoplasma from Cell Mask (cross-channel) in the model list and press Download. The Model Zoo verifies the checkpoint's SHA-256 after download, so a truncated or substituted file is rejected rather than silently used.

Model Zoo (Python)

python

from spacr import model_zoo

entry = next(e for e in model_zoo.catalogue() if e.key == "toxoplasma_from_cellmask_v1")
path  = model_zoo.install(entry, dest="~/spacr_models")
print(path)   # verified local checkpoint

Mask generation

Point spaCR's mask generation at the downloaded checkpoint:

python

from spacr.core import preprocess_generate_masks

settings = {
    "src": "/path/to/images",
    "pathogen": "cellpose",
    "pathogen_model": str(path),     # the checkpoint fetched above
    "pathogen_diameter": 12,
}
preprocess_generate_masks(settings)

In the GUI the same thing is under Make masks — choose the downloaded model in the Cellpose model field for the relevant object.

API: :func:spacr.core.preprocess_generate_masks, :func:spacr.spacr_cellpose.generate_masks_from_imgs

Performance

model	train	train obj.	test	test obj.	CV	F1 @ IoU 0.5	AJI	Dice	final train loss	final val loss	val - train	best epoch
stock cpsam_v2 (no fine-tuning)	—	—	463	6116	—	0.0215	0.0080	0.0201	—	—	—	—
this model	2567	31637	463	6116	no (single well-grouped split)	0.6058	0.4939	0.6096	0.0072	0.0099	+0.0028	100 / 100

Scored on a well-grouped held-out split — no well appears in both train and test — including fields with no parasites, so false positives are counted.

Per host cell line:

host	n	F1	AJI
HFF	159	0.5568	0.5649
HeLa	151	0.7105	0.5963
THP1	153	0.4652	0.3191

Objects are reference (ground-truth) objects. Training-set object counts were not recorded at training time; the held-out counts come from the scoring bundle.

Training curves

training curves

Loss is on a log scale. Train and validation tracking each other is the overfitting check: a validation curve that turns up while train keeps falling is the signature this model does not show.

Training data

2567 training fields and 463 held-out fields, split by well (training/split_by_well.csv) so no well leaks across the split. Targets are PV-regenerated masks (masks_pv). Hosts: HFF, HeLa and THP1.

Trained for 100 epochs from stock cpsam_v2, AdamW, lr 1e-5, weight decay 0.1, batch 1.

Provenance note. A power loss interrupted this run at 57/100 epochs. Training was continued from the epoch-50 checkpoint with the original learning-rate schedule replayed exactly from index 50 (validated bit-exactly against the interrupted run's recorded learning rates), so epochs 51-100 follow the schedule the uninterrupted run would have used. Cellpose stores net.state_dict() only, so the AdamW moments and augmentation RNG restarted; validation loss shows the two runs converged again within two epochs. Both epoch histories are in training/ for full transparency.

Environment


cellpose (training)	4.2.1.1
cellpose (scoring)	4.2.1.1
torch	2.13.0+cu130
python	3.12.13
GPU	NVIDIA GeForce RTX 3090
base weights	cpsam_v2

Files in this repository

path	what
`toxoplasma_from_cellmask_pv`	the checkpoint
`training/epoch_history.csv`	per-epoch losses + pixel metrics, epochs 51-100
`training/epoch_history_1-100.csv`	canonical per-epoch history, epochs 1-100 (1-50 from the interrupted run, 51-100 from the resume)
`training/epoch_history_epochs1-57_interrupted.csv`	the interrupted run, epochs 1-57
`training/loss_per_epoch.csv`	slim loss view
`training/run_config.json`	exact hyperparameters, versions, GPU
`training/split_by_well.csv`	the well-grouped split
`training/training_curves.png`	loss and checkpoint-metric curves
`evaluation/report.json`	full scorecard, all checkpoints, per host
`evaluation/metrics.csv`	F1/AJI/Dice per checkpoint and per host
`evaluation/best_perimage.csv`	per-image scores for the best checkpoint
`evaluation/comparison_vs_stock_summary.csv`	this model vs stock cpsam_v2
`evaluation/stock_cpsam_v2_summary.csv`	the stock baseline

Limitations

—The held-out split is used for checkpoint selection, so it is validation data rather than a fully independent test set.
—Targets are automatic reference labels (PV-regenerated masks), not hand-drawn ground truth.
—THP1 is the weakest host (F1 0.465); HeLa the strongest (0.711).
—Accuracy falls above IoU 0.8 — suited to counting, occupancy and area rather than precise morphometry.