{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Image segmentation with StarDist\n",
"\n",
"StarDist is a deep-learning based image segmentation method for segmenting objects such as cell nuclei.\n",
"\n",
"See also\n",
"* [StarDist in Proceedings of MICCAI 2018](https://link.springer.com/chapter/10.1007/978-3-030-00934-2_30)\n",
"* [StarDist on github](https://github.com/stardist/stardist)\n",
"* [StarDist NEUBIAS Academy webinar](https://www.youtube.com/watch?v=Amn_eHRGX5M)\n",
"* [StarDist prediction example](https://github.com/stardist/stardist/blob/master/examples/2D/3_prediction.ipynb)\n",
"\n",
"As usual, we start by loading an example image."
]
},
{
"cell_type": "code",
"execution_count": 10,
"metadata": {
"tags": []
},
"outputs": [],
"source": [
"from stardist.models import StarDist2D\n",
"from csbdeep.utils import normalize\n",
"from stardist import random_label_cmap\n",
"\n",
"import stackview\n",
"import matplotlib.pyplot as plt\n",
"import numpy as np\n",
"from skimage.data import human_mitosis"
]
},
{
"cell_type": "code",
"execution_count": 2,
"metadata": {
"tags": []
},
"outputs": [
{
"data": {
"text/html": [
"
\n",
"
\n",
"
\n",
"\n",
"
\n",
"
\n",
"\n",
"
\n",
"
shape
(512, 512)
\n",
"
dtype
uint8
\n",
"
size
256.0 kB
\n",
"
min
7
max
255
\n",
"
\n",
"\n",
"
\n",
"
\n",
"
"
],
"text/plain": [
"StackViewNDArray([[ 8, 8, 8, ..., 63, 78, 75],\n",
" [ 8, 8, 7, ..., 67, 71, 71],\n",
" [ 9, 8, 8, ..., 53, 64, 66],\n",
" ...,\n",
" [ 8, 9, 8, ..., 17, 24, 59],\n",
" [ 8, 8, 8, ..., 17, 22, 55],\n",
" [ 8, 8, 8, ..., 16, 18, 38]], dtype=uint8)"
]
},
"execution_count": 2,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"image = human_mitosis()\n",
"stackview.insight(image)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Loading a model\n",
"Models are files that typically contain a neural network which is capable of segmenting an image. StarDist comes with some pretrained models for demonstrating how the algorithm performs on a general use-case such as nuclei segmentation. If such a pretrained model does not perform well on your data (be a good scientist and check that carefully!), you need to [train your own model](https://github.com/stardist/stardist/blob/master/examples/2D/2_training.ipynb) which fits to your data. For training, you will likely need a powerful GPU, for applying a model (prediction step) to a 2D image no GPU is necessary."
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {
"tags": []
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Found model '2D_versatile_fluo' for 'StarDist2D'.\n",
"Loading network weights from 'weights_best.h5'.\n",
"Loading thresholds from 'thresholds.json'.\n",
"Using default values: prob_thresh=0.479071, nms_thresh=0.3.\n"
]
}
],
"source": [
"# creates a pretrained model\n",
"model = StarDist2D.from_pretrained('2D_versatile_fluo')"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Normalizing the input image\n",
"Many algorithms using neural networks need normalized input data to work on. For example, you can determine the 1% and the 99.8% percentile (that's very commond) and normalize your image so that the intensities spread between these percentiles are afterwards in the range between 0 and 1. We need to do this because the model was trained on an image in this range and might not be able to segment images with different intensity ranges."
]
},
{
"cell_type": "code",
"execution_count": 7,
"metadata": {
"tags": []
},
"outputs": [],
"source": [
"axis_norm = (0,1)\n",
"image = normalize(image, 1,99.8, axis=axis_norm)"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Segmenting the image and labeling the individual objects is often called \"instance segmentation\" or \"prediction\" in the artificial intelligence community."
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {
"tags": []
},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"1/1 [==============================] - 0s 127ms/step\n"
]
},
{
"data": {
"text/html": [
"