Cellpose-Planer

Cellpose is a generalist algorithm for cellular segmentation, Which written by Carsen Stringer and Marius Pachitariu.

Planer is a light-weight CNN framework implemented in pure Numpy-like interface. It can run only with Numpy. Or change different backends. (Cupy accelerated with CUDA, ClPy accelerated with OpenCL).

So Cellpose-Planer is the cellpose models on planer framework. We generate onnx from torch models, then deduce it to planer model.

We just use cellpose's models, but we rewrite all the pre-after processing and render algorithm, So the result is not same as the official one

Features

• cellpose-planer is very light, only depend on Numpy and Scipy.
• cellpose-planer can be accelerated with Cupy.
• without ui, with out object or class, pure function oriented designed.
• optimize cellpose 's pre-after processing and render algorithm, having a better performance and result.

Install

pip install cellpose-planer

Option: pip install cupy-cuda101 on envidia gpu, install cuda and cupy would get a large acceleration.

Usage

import cellpose_planer as cellpp
from skimage.data import coins

img = coins()
x = img.astype(np.float32)/255

net = cellpp.load_model('cyto_0')
flowpb, style = cellpp.get_flow(net, x, size=480)
lab = cellpp.flow2msk(flowpb, level=0.2)

flowpb = cellpp.asnumpy(flowpb)
lab = cellpp.asnumpy(lab)
cellpp.show(img, flowpb, lab)

first time: search and download models

search the models you need, and download them. (just one time)

>>> import cellpose_planer as cellpp
>>> cellpp.search_models()
cyto_0 : --
cyto_1 : --
cyto_2 : --
cyto_3 : --
nuclei_0 : --
nuclei_1 : --
nuclei_2 : --
nuclei_3 : --

>>> cellpp.download(['cyto_0', 'cyto_1', 'cyto_2', 'cyto_3'])
download cyto_0 from http://release.imagepy.org/cellpose-planer/cyto_0.npy
100%|█████████████████████████████████████| 100/100 [00:10<00:00,  2.37it/s]
download cyto_1 from http://release.imagepy.org/cellpose-planer/cyto_1.npy
100%|█████████████████████████████████████| 100/100 [00:10<00:00,  2.37it/s]
download cyto_2 from http://release.imagepy.org/cellpose-planer/cyto_2.npy
100%|█████████████████████████████████████| 100/100 [00:10<00:00,  2.37it/s]
download cyto_3 from http://release.imagepy.org/cellpose-planer/cyto_3.npy
100%|█████████████████████████████████████| 100/100 [00:10<00:00,  2.37it/s]

>>> cellpp.list_models()
['cyto_0', 'cyto_1', 'cyto_2', 'cyto_3']

1. load models

you can load one model or more, when multi models, you would get a mean output.

nets = cellpp.load_model('cyto_0')
nets = cellpp.load_model(['cyto_0', 'cyto_1', 'cyto_2', 'cyto_3'])

2. get flow image

def get_flow(nets, img, cn=[0,0], sample=1, size=512, tile=True, work=1, callback=progress)

• nets: the nets loaded upon.

• img: the image to process

• cn: the cytoplasm and nucleus channels

• sample: if not 1, we scale it. (only avalible when tile==True)

• size: when tile==True, this is the tile size, when tile==False, we scale the image to size.

• tile: if True, method try to process image in tiles. else resize the image.

• work: open multi-thread to process the image. (GPU not recommend)

flowpb, style = cellpp.get_flow(net, coins(), [0,0], work=4)

3. flow to mask

def flow2msk(flowpb, level=0.5, grad=0.5, area=None, volume=None)

• flowpb: get_flow 's output

• level: below level means background, where water can not flow. So level decide the outline.

• grad: if the flow gradient is smaller than this value, we set it 0. became a watershed. bigger gradient threshold could suppress the over-segmentation. especially in narrow-long area.

• area: at end of the flow process, every watershed should be small enough. (<area), default is 0 (auto).

• volume: and in small area, must contian a lot of water. (>volume), default is 0 (auto).

msk = cellpp.flow2msk(flowpb, level=0.5, grad=0.5, area=None, volume=None)

4. render

cellpose-planer implements some render styles.

import cellpose_planer as cellpp

# get edge from label msask
edge = cellpp.msk2edge(lab)
# get build flow as hsv 2 rgb
hsv = cellpp.flow2hsv(flow)
# 5 colors render (different in neighborhood)
rgb = cellpp.rgb_mask(img, lab)
# draw edge as red line
line = cellpp.draw_edge(img, lab)

5. backend and performance

Planer can run with numpy or cupy backend, by default, cellpose-planer try to use cupy backend, if failed, use numpy backend. But we can change the backend manually. (if you switch backend, the net loaded befor would be useless, reload them pleanse)

import cellpose-planer as cellpp

# use numpy and scipy as backend
import numpy as np
import scipy.ndimage as ndimg
cellpp.engine(np, ndimg)

# use cupy and cupy.scipy as backend
import cupy as cp
import cupyx.scipy.ndimage as cpimg
cellpp.engine(cp, cpimg)

here we time a 1024x1024 image on I7 CPU and 2070 GPU.

user switch engine: numpy
    net cost: 11.590
    flow cost: 0.0797

user switch engine: cupy
    net cost: 0.0139
    flow cost: 0.009

Model deducing and releasing

Planer only has forward, so we need train the models in torch. then deduc it in planer.

deduce from torch

# train in cellpose with torch, and export torch as onnx file.
from planer import onnx2planer
onnx2planer(xxx.onnx)

then you would get a json file (graph structure), and a npy file (weights).

model releasing

if you want to share your model in cellpose-planer, just upload the json and npy file generated upon to any public container, then append a record in the models list tabel below, and give a pull request. infact, when we call cellpp.search_models, cellpp pull the text below and parse them.

models list

model name	auther	description	url
cyto_0	carsen-stringer	for cell cyto segmentation	download
cyto_1	carsen-stringer	for cell cyto segmentation	download
cyto_2	carsen-stringer	for cell cyto segmentation	download
cyto_3	carsen-stringer	for cell cyto segmentation	download
nuclei_0	carsen-stringer	for cell nuclear segmentation	download
nuclei_1	carsen-stringer	for cell nuclear segmentation	download
nuclei_2	carsen-stringer	for cell nuclear segmentation	download
nuclei_3	carsen-stringer	for cell nuclear segmentation	download

cellpp.search_models function pull the text below and parse them, welcom to release your models here!

Use cellpose-planer as ImagePy plugins

Path: https://gitee.com/imagepy/cellpose-planer

Version: 0.1

Author: YXDragon, Y.Dong

Email: yxdragon@imagepy.org

Keyword: cellpose, segment, unet

Description: cellpose on planer framework for imagepy.

cellpose-planer can also start as ImagePy's plugins. supporting interactive and bat processing.