LACSS is a deep-learning model for 2D/3D single-cell segmentation from microscopy images.
pip install lacss| name | #params | download | mAP LiveCell* | mAP Cellpose* | mAP NIPS* | ovule (3D)* | platynereis (3D)* |
|---|---|---|---|---|---|---|---|
| small | 60M | model | 56.3 | 52.0 | 54.2 | 44.4 | 56.7 |
| base | 152M | model | 57.1 | 56.0 | 62.9 | 47.0 | 60.8 |
- mAP is the average of APs at IOU threshoulds of 0.5-0.95 (10 segments). Evaluations are on either testing or validation split of the corresponding datasets.
| name | #params | training data | download | AP50 | AP75 | mAP |
|---|---|---|---|---|---|---|
| small-2dL | 40M | LiveCell | model | 84.3 | 61.1 | 57.4 |
| small-2dC | 40M | Cellpose+Cyto2 | model | 87.6 | 62.0 | 56.4 |
| small-2dN | 40M | NIPS challenge | model | 84.6 | 64.8 | 57.3 |
You can now deploy the pretrain models as GRPC server:
python -m lacss.deploy.remote_server --modelpath=<model_file_path>For a GUI client see the Trackmate-Lacss project, which provides a FIJI/ImageJ plugin to perform cell segmentation/tracking in an interactive manner.
-
multi-modality: works on both 2D (multichannel) images and 3D image stacks.
-
Speed: Inference time of the base model (150M parameters) is < 200 ms on GPU for an 1024x1024x3 image. We achieve this by desigining an end-to-end algorithm and aggressively eliminate CPU-dependent post-processings.
-
Point-supervised traing: Lacss is a multi-task model with a separate "localization" head (beside the segmentation head) predicting cell locations. This also means that you can train/fine-tune cell-segmentation models using only point labels. See refernces for details.
