ounissimehdi / point2cell

Point2Cell:

Efficient interactive biomedical image annotation. Instantiation to stain-free phase-contrast microscopy

🛠️ Installation Steps

1. Install anaconda-python (for more details please check https://docs.anaconda.com/anaconda/install/index.html)

2. Clone the repository (this will take some time)

git clone https://github.com/ounissimehdi/Point2Cell

3. Change the working directory

cd Point2Cell

4. Create the point2Cell conda environment with all the dependencies

conda env create -f environment.yml

Note that this project is tested on: Windows: 11, MacOS: BigSur and Lunix: Ubuntu 20.04.3 LTS.

with the last version of Pytorch ( pytorch 1.10.2:cuda 11.3 and cudnn 8.0 ) to this date.

It supports GPU 😊

5. Activate the point2Cell conda environment

conda activate point2Cell

🌟 You are all set!

🥽 How to use Point2Cell -from scratch-:

It is highly recommended to read the paper before getting to use the Point2Cell annotation tool (please do not forget to cite the paper and the github 😊).

In this tutorial, a public dataset will be used, it is called "HeLa cells on a flat glass", it can be downloaded by using the following link (http://data.celltrackingchallenge.net/training-datasets/DIC-C2DH-HeLa.zip).

Some basic information about this dataset :

• Microscope: Zeiss LSM 510 Meta

• Objective lens: Plan-Apochromat 63x/1.4 (oil)

• Pixel size (microns): 0.19 x 0.19

• Time step (min): 10

Make sure you are at the same folder level as the repository Point2Cell/

1. Create the dataset folder and prepare the dataset

mkdir dataset
cd dataset

mkdir hela_cells_dataset
cd hela_cells_dataset

wget http://data.celltrackingchallenge.net/training-datasets/DIC-C2DH-HeLa.zip
unzip DIC-C2DH-HeLa

rm DIC-C2DH-HeLa.zip

cd ../../

Note that the folder (dataset\hela_cells_dataset\DIC-C2DH-HeLa\01) with 84 images will be used for training/validation, and the folder (dataset\hela_cells_dataset\DIC-C2DH-HeLa\02) with another 84 images will be used for testing.

At the end of this step you should have:

tree
YOUR_PATH_TO_MAIN_FOLDER:.
├───dataset
│   └───hela_cells_dataset
│       └───DIC-C2DH-HeLa
│           ├───01
│           ├───01_GT
│           │   ├───SEG
│           │   └───TRA
│           ├───01_ST
│           │   └───SEG
│           ├───02
│           ├───02_GT
│           │   ├───SEG
│           │   └───TRA
│           └───02_ST
│               └───SEG
└───Point2Cell
    ├───annotation_interface
    ├───demo
    ├───experiments
    ├───labelme_utils
    ├───train_models
    │   └───hela_cells
    │       └───runs
    ├───unet
    └───utils

2. Prepare HeLa cells dataset train/validation/test splits after data-augmentation

cd Point2Cell/
jupyter-notebook

a. Please open hela_data_preparation.ipynb inside the Point2Cell, your browser should open with a link similar to this one (http://localhost:8888/notebooks/hela_data_preparation.ipynb).

b. Execute all cells on the notebook, this notebook will apply data-augmentation to the training/validation data.

At the end of this step you should have:

tree
YOUR_PATH_TO_MAIN_FOLDER:.
├───dataset
│   ├───augmented_hela_cells_dataset
│   │   ├───test
│   │   ├───train
│   │   └───val
│   └───hela_cells_dataset
│       └───DIC-C2DH-HeLa
│           ├───01
│           ├───01_GT
│           │   ├───SEG
│           │   └───TRA
│           ├───01_ST
│           │   └───SEG
│           ├───02
│           ├───02_GT
│           │   ├───SEG
│           │   └───TRA
│           └───02_ST
│               └───SEG
└───Point2Cell
    ├───annotation_interface
    ├───demo
    ├───experiments
    ├───labelme_utils
    ├───train_models
    │   └───hela_cells
    │       └───runs
    ├───unet
    └───utils

Since tha data is ready!

• All train/validation dataset images : 84 (672 images with data-augmentation)

  • (80%) train images : 67 (536 images with data-augmentation)
  • (20%) validation images : 17 (136 images with data-augmentation)

• Test dataset contain 84 images and their ground truth masks.

Now we can start training the two U-Net models ('intra cellular density' and 'cell mask predictor').

3. rain/validate then test the U-Net -cell mask predictor- model

cd Point2Cell/train_models/hela_cells/
conda activate point2Cell
python bn_mask_train_unet.py   

4. Train/validate then test the U-Net -intra cellular density- model

cd Point2Cell/train_models/hela_cells/
conda activate point2Cell
python density_mask_train_unet.py   

5. Check the training/validation and testing details inside (./Point2Cell/experiments/YOUR_EXP_NAME/logfile.log)

Note that this project support the TensorBoard, all records can be found inside (./Point2Cell/train_models/hela_cells/runs).

cd Point2Cell/train_models/hela_cells/
tensorboard --logdir runs

Happy visualization 😎!!

6. Compute density maps and cell masks

cd Point2Cell/annotation_interface/
jupyter-notebook  

a. Please open novel_data_prep.ipynb inside the (./Point2Cell/annotation_interface/) folder, your browser should open with a link similar to this one (http://localhost:8888/notebooks/novel_data_prep.ipynb).

b. Execute all cells on the notebook, this notebook will apply load the best models already trained and use them to compute the density maps and cell masks for the novel data (test dataset).

At the end of this step you should have:

tree
YOUR_PATH_TO_MAIN_FOLDER:.
C:.
├───dataset
│   ├───augmented_hela_cells_dataset
│   │   ├───test
│   │   ├───train
│   │   └───val
│   └───hela_cells_dataset
│       ├───DIC-C2DH-HeLa
│       │   ├───01
│       │   ├───01_GT
│       │   │   ├───SEG
│       │   │   └───TRA
│       │   ├───01_ST
│       │   │   └───SEG
│       │   ├───02
│       │   ├───02_GT
│       │   │   ├───SEG
│       │   │   └───TRA
│       │   └───02_ST
│       │       └───SEG
│       └───novel_annotation
│           ├───binary_masks
│           ├───density_masks
│           └───images
└───Point2Cell
    ├───annotation_interface
    ├───demo
    ├───experiments
    ├───labelme_utils
    ├───train_models
    │   └───hela_cells
    │       └───runs
    ├───unet
    └───utils

7. Start the annotation of the novel dataset

cd Point2Cell/annotation_interface/
python annotation_tool.py 

Enjoy😊!

🔬 Quantitative results:

The best model is based on the smallest validation loss.

HeLa dataset: Cell mask predictor (binary masks):

data-split/metrics	Mean BCE Loss	DICE	PRECISION	RECALL	ACCURACY
Test	0.151±0.0121	93.92%±0.5107	96.30%±0.4860	91.77%±1.3294	93.81%±0.4777
Validation	0.1237±0.0026	94.88%±0.0606	93.82%±0.4604	96.02%±0.5218	94.85%±0.0798

Detailed results on the test dataset:

Cross-validation/metrics	Mean BCE Loss	DICE	PRECISION	RECALL	ACCURACY
FOLD 1	0.1644	93.40%	97.14%	90.00%	93.37%
FOLD 2	0.1414	94.36%	96.05%	92.90%	94.23%
FOLD 3	0.1348	94.58%	96.09%	93.18%	94.43%
FOLD 4	0.1493	94.01%	95.73%	92.42%	93.86%
FOLD 5	0.1651	93.29%	96.51%	90.35%	93.19%

Detailed results on the validation:

Cross-validation/metrics	Mean BCE Loss	DICE	PRECISION	RECALL	ACCURACY
FOLD 1	0.1260	94.86%	94.29%	95.48%	94.79%
FOLD 2	0.1219	94.93%	93.49%	96.50%	94.86%
FOLD 3	0.1252	94.93%	93.26%	96.72%	94.83%
FOLD 4	0.1263	94.77%	93.64%	95.98%	94.80%
FOLD 5	0.1195	94.91%	94.44%	95.43%	95.01%

HeLa dataset: Intra-cellular density predictor:

data-split/metrics	Mean MSE
Test	0.03164±0.0028
Validation	0.0252±0.0004

Detailed results on the test dataset:

Cross-validation/metrics	Mean MSE
FOLD 1	0.0338
FOLD 2	0.0295
FOLD 3	0.0274
FOLD 4	0.0350
FOLD 5	0.0325

Detailed results on the validation:

Cross-validation/metrics	Mean MSE
FOLD 1	0.0256
FOLD 2	0.0258
FOLD 3	0.0252
FOLD 4	0.0249
FOLD 5	0.0246

🆚 Point2Cell (our tool) VS LabelMe (polygon based annotation tool):

Qualitative results:

Some examples annotated by LabelMe and Point2Cell respectively, compared to the ground truth (DICE coefficient percentage).

Annotation speed comparison between Point2Cell and LabelMe:

Quantitative results:

	Time per image	Average Dice coefficient
Point2Cell (our tool)	14.6 sec (std 1.35sec)	94.97% (std 0.88%)
LabelMe	96.1 sec (std 8.03sec)	91.30% (std 0.90%)

Detailed results:

	LabelMe		Point2Cell
images/metrics	Time (sec)	DICE	Time (sec)	DICE
img_0	108	92.05%	13	95.96%
img_1	95	92.81%	13	95.68%
img_2	108	91.52%	15	95.14%
img_3	87	90.2%	14	94.31%
img_4	85	91.32%	15	95.59%
img_5	88	90.69%	14	93.64%
img_6	102	91.66%	16	94.08%
img_7	99	89.7%	17	93.71%
img_8	90	92.2%	16	95.76%
img_9	99	90.94%	13	95.87%

email address : [email protected]

@misc{Github,
  author={Mehdi Ounissi and Daniel Racoceanu},
  title={Point2Cell: efficient learning for interactive biomedical image annotation. Instantiation to stain-free phase-contrast microscopy},
  year={2022},
  url={https://github.com/ounissimehdi/Point2Cell},
}