repository·main·Indexed 41 days ago
https://github.com/fzj-inm1-bda/celldetectionA Python library for cell segmentation and detection in biological imaging. It provides pretrained U-Net architectures, Contour Proposal Networks (CPN), and utility functions for image processing and visualization. The library supports 2D and 3D architectures including MA-Nets, ConvNeXt, and Residual Networks, with integrations for timm and segmentation_models_pytorch (smp). It includes a Napari plugin for interactive visualization and supports deployment via Docker, Apptainer, and Hugging Face API.
The celldetection.models module provides a variety of neural network architectures designed for cell detection tasks. These include specialized segmentation networks and various encoder backbones.
Available architectures include:
Timm Encoder and Smp Encoder to integrate pre-trained weights.The celldetection.data.datasets module provides access to various datasets used for cell detection tasks. Supported datasets include:
Use these modules to load standardized data formats for training or evaluation.
You can leverage the timm (PyTorch Image Models) library to use high-quality pretrained encoders within celldetection. This allows you to build models like CpnTimmMaNet or TimmUNet. You can also use cd.models.TimmEncoder to wrap a timm model as an encoder for other architectures.
import timm
import celldetection as cd
# Explore available timm models
timm.list_models(filter='*')
# Use a timm-based model
# Examples: cd.models.CpnTimmMaNet, cd.models.TimmUNet, cd.models.TimmMaNetFor specialized analysis and performance measurement, use the following modules:
celldetection.data.cpn): Operations related to CPN (likely Cell/Point/Nuclei) processing.celldetection.data.instance_eval): Tools for instance-level evaluation, used to calculate metrics for cell detection accuracy.The --tile_size and --stride arguments control how large images are partitioned for processing.
--tile_size: The dimension of the processing window.--stride: The step size between windows.A common strategy for balancing computational efficiency and stitching quality is to set the stride to a value slightly less than the tile size (e.g., tile_size=1024 and stride=768).
The celldetection.data package includes modules for common data processing tasks:
celldetection.data.misc): General utility functions for data handling.celldetection.data.segmentation): Tools specifically designed for processing segmentation masks and related data structures.The project integrates with segmentation_models_pytorch (smp). You can use cd.models.SmpEncoder to create an encoder from an smp backbone (e.g., 'mit_b5') with pretrained weights. This encoder can then be used in architectures like CpnSmpMaNet, CpnSmpUNet, SmpMaNet, or SmpUNet.
import segmentation_models_pytorch as smp
import celldetection as cd
# Explore available smp encoders
smp.encoders.get_encoder_names()
# Create an smp encoder for use in celldetection
encoder = cd.models.SmpEncoder(encoder_name='mit_b5', pretrained='imagenet')
# Use smp-based models like cd.models.CpnSmpUNet or cd.models.SmpUNetIf you encounter memory issues while running cpn_inference.py, adjust the following parameters to reduce the memory footprint:
--tile_size: Defines the size of the windows used to process large images. Reducing this (e.g., to 512) consumes less memory but may increase processing time or affect segmentation quality for large objects. (Default: 1024)--stride: Determines the overlap between consecutive tiles. A smaller stride increases overlap (better stitching) but increases computation. A larger stride reduces computation but may miss objects at tile borders. (Default: 768)--batch_size: Use a smaller batch size to reduce the number of samples processed simultaneously.--precision: Lower the precision (e.g., from '32-true' to '16-mixed') to reduce the model's memory footprint, though this may impact accuracy.--strategy and --devices settings to distribute the workload more effectively.For High-Performance Computing (HPC) environments using Slurm and Apptainer, use srun to execute the script within a Singularity/Apptainer image.
srun [srun-arguments] apptainer exec --nv celldetection_latest.sif python cpn_inference.py [arguments]srun [srun-arguments] apptainer exec --nv celldetection_latest.sif python cpn_inference.py [arguments]You can install the celldetection package using pip. Ensure that PyTorch is already installed on your system before proceeding.
# Install from PyPI
pip install -U celldetection
# Install directly from GitHub
pip install git+https://github.com/FZJ-INM1-BDA/celldetection.gitcelldetection.data.toydata module provides small, simplified datasets. These are intended for rapid prototyping, debugging, and verifying that your data pipelines or model integrations are working correctly without the overhead of large biological datasets.