ml-project-2-cellseg

Project 2 for ML_course : Cell Segmentation in 3D

Authors : Cyril Achard Yves Paychère Colin Hoffman

Important

This is based on previous code from the Mathis Lab of Adaptive Motor Control, made by Cyril Achard and Maxime Vidal. Authors and updates are indicated in the files' header comment as precisely as possible.

Installation procedure

From the Drive (link is in the report) :

Copy the dataset_clean folder to obtain the training and validation datasets.
Weights can be found the weights_results folder and can be used for inference. The dataset_cleanfolder is also needed for inference.

We recommend using anaconda to install the required packages in an env. The following commands will create a new conda environment named cellseg and install required packages.

conda create --name cellseg python=3.8
conda activate cellseg
Install pytorch according to the instructions on the website, with python 3.8 and using pip. Use cuda if relevant, it is strongly recommended even for inference.
pip install -r requirements.txt

Guided overview

In order to do obtain the result for the unsupervised model you should follow these steps. Note that all the notebooks and python files for this guide are on models/wnet/:

You should have downloaded the folder dataset_clean and added it in the main folder, You should download the file weights_results/results_unsupervised/new_results_wnet_1500e.zip from the Drive and extract its content into a folder that you should create weights_results/results_unsupervised/
If you want to do the inference (not recommended as it takes multiple Go of ram) You should run the notebook predict.ipynb and run the last cell for each image segmentated. To do so you should select the different layers 'Segmentation 1 image' and run the cell afterward. Then you can save the new layers directly from napari
You can directly download the results that are on the folder results (output and statistics)/self supervised/ and put them on the folder results/self supervised/
To relabel the image used to evaluate the performance of the unsupervised model, you should run the script relabel_images.py. This will open a window showing the previously labeled neurons that were separated by the script. Once the script is finished, you can close the window and see the neurons that were previously missed. To view the image in 3D, you can click on the button with a square in the bottom left corner of the napari window, or press Ctrl+Y. You can change the view by clicking on the image and moving your mouse. On the left panel, you should have three layers: labels, potential neurons, and images. To see only the potential neurons on top of the image, you should click on the eye button of the labels layer. You will see two neurons. To find their values, you should click on the "potential neurons" layer and move your mouse over the neurons. The bottom left corner will show the label values, such as "potential neurons [21 19 9]: 13", which means that one of the label values you can add is 13. In the terminal where you run the code, you will see the prompt "Which labels do you want to add (0 to skip)? (separated by a comma):". To add the two forgotten neurons, you should enter "13,29" and press enter. Then, close the napari window and answer "n" to the two following questions. This will save the new labels.
To evaluate the model performances on this image you can run the notebook evaluate_performances.ipynb. The image used in the report is the 2nd one
To visualize the evaluation metric you can run the file visualize_performances.py. You can compare the classification of the labels between found, not found, artifacts and fused by hiding all the other layers (by clicking on the eye button) and comparing what the model found and what the ground truth is.

Quality check and relabelling

All the scripts and notebooks that are used to label the images and perform a quality check of the labels are in the folder label. To relabel the labels to ensure that they are unique, check their quality and add neurons missed you can use the function relabel in the file relabel.py, as shown in the main of this file. To label new images with a basic watershed algorithm you can use the function make_labels of the file make_artefact_labels. To label all the artefacts of the dataset you can run the main of the file make_artefact_labels.py.

Data augmentation

In order to add artefacts to the images labelled you can use the function add_artefacts_to_folder of the file data_augmentation in the folder data_augmentation, as shown in the notebook data_augmentation.ipynb in the same folder.

Training models on axons and augmented data

You can use train_multichannel.py to re-train models to attempt to reproduce results. Use the dataset_clean folder on the Drive, place it in the repo path, and follow the instructions in the main of the file to reproduce results.

Inference and evaluation

Use infer_and_evaluate.py to perform inference on validation data and evaluate the results. The weights are available in the weights_results folder on the Drive. The dataset_clean folder is also needed for inference.

Unsupervised WNet model

The unsupervised WNet model is in the folder models/wnet. To train the model you can use train_wnet.py by either executing it or calling the method train from a notebook. To perform inference you can use the PredictWNet class from predict.py in a notebook. predict.ipynb is such a notebook and can also be used to visualize the results in napari. The weights of the model are available in the folder on the Drive. Use config.py to change the parameters of the model. For the time being, please don't use the CRF post-processing as it is doesn't work due to an implementation error from the library used.

image processing

If you want to obtain the results presented in the report for the image processing you should use the notebook model_watershed.ipynb that is in the main folder

Libraries used

MONAI
PyTorch
pydensecrf
cython
napari
scikit-image
tifffile
Others (see requirements.txt)

C-Achard / ml-project-2-cellseg