Volume Electron Microscopy SEGmentation

A package for the segmentation of volume electron microscopy images using XGBoost.

Installation

You can install vemseg via pip:

pip install vemseg

Usage

Shown below are some example uses of the VEMClassifier and some other functionality provided by the vemseg package.

import vemseg as vem
from skimage.io import imread

X = imread("train_image.tiff")
y = imread("train_labels.tiff")
mask = imread('train_mask.tiff')

clf = vem.VEMClassifier(max_depth=4,
                         n_estimators=400,
                         learning_rate=0.01,
                         verbosity=0)

features = clf.set_features(
    ["sobel_of_gaussian_blur","laplace_box_of_gaussian_blur", "original"], 
    range(1, 4)
)

clf.fit(X, y, mask=mask)

We then use clf.predict to get the prediction from the trained classifier. We can use the show flag to generate an image of the predictions.

prediction = clf.predict(X, show=True, mask=mask)

We can then plot the features importance like so, with fig size indicating the size of the plot.

clf.feature_importance(fig_size=(20, 10))

Deciding what hyperparameters to use can often be very difficult, so we can use an adaptation of sklearns GridSearchCV to determine which hyperparameters work best.

parameters = {
    'max_depth': range(2, 5, 1),
    'n_estimators': range(200, 500, 100)
}

grid_search = clf.GridSearchCV(
    X,
    y,
    param_grid=parameters,
    n_jobs=10,
    cv=10,
    verbose=True,
    update_params=True,
    mask=mask,
)

print(grid_search.best_estimator_)

Will output

Fitting 10 folds for each of 9 candidates, totalling 90 fits
VEMClassifier(base_score=0.5, booster='gbtree', callbacks=None,
              colsample_bylevel=1, colsample_bynode=1, colsample_bytree=1,
              early_stopping_rounds=None, enable_categorical=False,
              eval_metric=None, feature_types=None, gamma=0, gpu_id=-1,
              grow_policy='depthwise', importance_type=None,
              interaction_constraints='', learning_rate=0.300000012,
              max_bin=256, max_cat_threshold=64, max_cat_to_onehot=4,
              max_delta_step=0, max_depth=3, max_leaves=0, min_child_weight=1,
              missing=nan, monotone_constraints='()', n_estimators=200,
              n_jobs=0, num_parallel_tree=1, objective='binary:logistic',
              predictor='auto', ...)

After completing training one can save the model to be reused like so.

clf.save_model('model.json')

And to reload the model from the JSON file one simply has to initialise a new VEMClassifier and then load the model.

import vemseg

pre_trained_model = vem.VEMClassifier()
pre_trained_model.load_model('model.json')

License

Distributed under the terms of the BSD-3 license, "napari-vemseg" is free and open source software

Contribute to VEMseg

Contributions are very welcome.

Similar Software

Full credit to the following software for inspiration and use within this tool.

• apoc
• XGBoost
• scikit-learn

Reference