Skip to content

Latest commit

 

History

History
476 lines (354 loc) · 24.1 KB

README.md

File metadata and controls

476 lines (354 loc) · 24.1 KB
Raw

Cell Detection

Downloads Test PyPI Documentation Status DOI

⭐ Showcase

NeurIPS 22 Cell Segmentation Competition

neurips22 https://openreview.net/forum?id=YtgRjBw-7GJ

Nuclei of U2OS cells in a chemical screen

bbbc039 https://bbbc.broadinstitute.org/BBBC039 (CC0)

P. vivax (malaria) infected human blood

bbbc041 https://bbbc.broadinstitute.org/BBBC041 (CC BY-NC-SA 3.0)

🛠 Install

Make sure you have PyTorch installed.

PyPI

pip install -U celldetection

GitHub

pip install git+https://github.com/FZJ-INM1-BDA/celldetection.git

💾 Trained models

model = cd.fetch_model(model_name, check_hash=True)
model name training data link
ginoro_CpnResNeXt101UNet-fbe875f1a3e5ce2c BBBC039, BBBC038, Omnipose, Cellpose, Sartorius - Cell Instance Segmentation, Livecell, NeurIPS 22 CellSeg Challenge 🔗
Run a demo with a pretrained model 
import torch, cv2, celldetection as cd
from skimage.data import coins
from matplotlib import pyplot as plt

# Load pretrained model
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = cd.fetch_model('ginoro_CpnResNeXt101UNet-fbe875f1a3e5ce2c', check_hash=True).to(device)
model.eval()

# Load input
img = coins()
img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB)
print(img.dtype, img.shape, (img.min(), img.max()))

# Run model
with torch.no_grad():
    x = cd.to_tensor(img, transpose=True, device=device, dtype=torch.float32)
    x = x / 255  # ensure 0..1 range
    x = x[None]  # add batch dimension: Tensor[3, h, w] -> Tensor[1, 3, h, w]
    y = model(x)

# Show results for each batch item
contours = y['contours']
for n in range(len(x)):
    cd.imshow_row(x[n], x[n], figsize=(16, 9), titles=('input', 'contours'))
    cd.plot_contours(contours[n])
    plt.show()

🔬 Architectures

import celldetection as cd
Contour Proposal Networks
PyTorch Image Models (timm)

Also have a look at Timm Documentation.

import timm

timm.list_models(filter='*')  # explore available models
Segmentation Models PyTorch (smp) 
import segmentation_models_pytorch as smp

smp.encoders.get_encoder_names()  # explore available models
encoder = cd.models.SmpEncoder(encoder_name='mit_b5', pretrained='imagenet')

Find a list of Smp Encoders in the smp documentation.

U-Nets 
# U-Nets are available in 2D and 3D
import celldetection as cd

model = cd.models.ResNeXt50UNet(in_channels=3, out_channels=1, nd=3)
MA-Nets 
# Many MA-Nets are available in 2D and 3D
import celldetection as cd

encoder = cd.models.ConvNeXtSmall(in_channels=3, nd=3)
model = cd.models.MaNet(encoder, out_channels=1, nd=3)
Feature Pyramid Networks
ConvNeXt Networks 
# ConvNeXt Networks are available in 2D and 3D
import celldetection as cd

model = cd.models.ConvNeXtSmall(in_channels=3, nd=3)
Residual Networks 
# Residual Networks are available in 2D and 3D
import celldetection as cd

model = cd.models.ResNet50(in_channels=3, nd=3)
Mobile Networks

🐳 Docker

Find us on Docker Hub: https://hub.docker.com/r/ericup/celldetection

You can pull the latest version of celldetection via:

docker pull ericup/celldetection:latest
CPN inference via Docker with GPU 
docker run --rm \
  -v $PWD/docker/outputs:/outputs/ \
  -v $PWD/docker/inputs/:/inputs/ \
  -v $PWD/docker/models/:/models/ \
  --gpus="device=0" \
  celldetection:latest /bin/bash -c \
  "python cpn_inference.py --tile_size=1024 --stride=768 --precision=32-true"
CPN inference via Docker with CPU 
docker run --rm \
  -v $PWD/docker/outputs:/outputs/ \
  -v $PWD/docker/inputs/:/inputs/ \
  -v $PWD/docker/models/:/models/ \
  celldetection:latest /bin/bash -c \
  "python cpn_inference.py --tile_size=1024 --stride=768 --precision=32-true --accelerator=cpu"

Apptainer

You can also pull our Docker images for the use with Apptainer (formerly Singularity) with this command:

apptainer pull --dir . --disable-cache docker://ericup/celldetection:latest

🤗 Hugging Face Spaces

Find us on Hugging Face and upload your own images for segmentation: https://huggingface.co/spaces/ericup/celldetection

There's also an API (Python & JavaScript), allowing you to utilize community GPUs (currently Nvidia A100) remotely!

Hugging Face API

Python

from gradio_client import Client

# Define inputs (local filename or URL)
inputs = 'https://raw.githubusercontent.com/scikit-image/scikit-image/main/skimage/data/coins.png'

# Set up client
client = Client("ericup/celldetection")

# Predict
overlay_filename, img_filename, h5_filename, csv_filename = client.predict(
    inputs,  # str: Local filepath or URL of your input image
    
    # Model name
    'ginoro_CpnResNeXt101UNet-fbe875f1a3e5ce2c',
    
    # Custom Score Threshold (numeric value between 0 and 1)
    False, .9,  # bool: Whether to use custom setting; float: Custom setting
    
    # Custom NMS Threshold
    False, .3142,  # bool: Whether to use custom setting; float: Custom setting
    
    # Custom Number of Sample Points
    False, 128,  # bool: Whether to use custom setting; int: Custom setting
    
    # Overlapping objects
    True,  # bool: Whether to allow overlapping objects
    
    # API name (keep as is)
    api_name="/predict"
)


# Example usage: Code below only shows how to use the results
from matplotlib import pyplot as plt
import celldetection as cd
import pandas as pd

# Read results from local temporary files
img = imread(img_filename)
overlay = imread(overlay_filename)  # random colors per instance; transparent overlap
properties = pd.read_csv(csv_filename)
contours, scores, label_image = cd.from_h5(h5_filename, 'contours', 'scores', 'labels')

# Optionally display overlay
cd.imshow_row(img, img, figsize=(16, 9))
cd.imshow(overlay)
plt.show()

# Optionally display contours with text
cd.imshow_row(img, img, figsize=(16, 9))
cd.plot_contours(contours, texts=['score: %d%%\narea: %d' % s for s in zip((scores * 100).round(), properties.area)])
plt.show()

Javascript

import { client } from "@gradio/client";

const response_0 = await fetch("https://raw.githubusercontent.com/scikit-image/scikit-image/main/skimage/data/coins.png");
const exampleImage = await response_0.blob();
						
const app = await client("ericup/celldetection");
const result = await app.predict("/predict", [
    exampleImage,  // blob: Your input image
    
    // Model name (hosted model or URL)
    "ginoro_CpnResNeXt101UNet-fbe875f1a3e5ce2c",
    
    // Custom Score Threshold (numeric value between 0 and 1)
    false, .9,  // bool: Whether to use custom setting; float: Custom setting
    
    // Custom NMS Threshold
    false, .3142,  // bool: Whether to use custom setting; float: Custom setting
    
    // Custom Number of Sample Points
    false, 128,  // bool: Whether to use custom setting; int: Custom setting
    
    // Overlapping objects
    true,  // bool: Whether to allow overlapping objects
    
    // API name (keep as is)
    api_name="/predict"
]);

🧑‍💻 Napari Plugin

Find our Napari Plugin here: https://github.com/FZJ-INM1-BDA/celldetection-napari
Find out more about Napari here: https://napari.org bbbc039 You can install it via pip:

pip install git+https://github.com/FZJ-INM1-BDA/celldetection-napari.git

🏆 Awards

📝 Citing

If you find this work useful, please consider giving a star ⭐️ and citation:

@article{UPSCHULTE2022102371,
    title = {Contour proposal networks for biomedical instance segmentation},
    journal = {Medical Image Analysis},
    volume = {77},
    pages = {102371},
    year = {2022},
    issn = {1361-8415},
    doi = {https://doi.org/10.1016/j.media.2022.102371},
    url = {https://www.sciencedirect.com/science/article/pii/S136184152200024X},
    author = {Eric Upschulte and Stefan Harmeling and Katrin Amunts and Timo Dickscheid},
    keywords = {Cell detection, Cell segmentation, Object detection, CPN},
}

🔗 Links

🧑‍🔬 Thanks!

Stargazers repo roster for @FZJ-INM1-BDA/celldetection Forkers repo roster for @FZJ-INM1-BDA/celldetection