Learned Segmenter¶

This example trains a segmenter from a volume annotation and applies it to the whole dataset.

Result of the trained segmenter on the full dataset

It builds upon the two previous examples using the Dataset API and dataset upload. Additionally, it downloads this manual volume annotation of a subset of the skin example dataset which is used for training.

This example additionally needs the scikit-learn and pooch packages.

import os
from functools import partial
from tempfile import TemporaryDirectory
from time import gmtime, strftime

import numpy as np
from skimage import feature
from skimage.future import TrainableSegmenter

import webknossos as wk


def main() -> None:
    # We are going to use a public demo annotation for this example

    annotation = wk.Annotation.download(
        "https://webknossos.org/annotations/616457c2010000870032ced4"
    )

    # Step 1: Read the training data from the annotation and the dataset's color
    # layer (the data will be streamed from WEBKNOSSOS to our local computer)
    training_data_bbox = annotation.user_bounding_boxes[0]  # type: ignore[index]
    time_str = strftime("%Y-%m-%d_%H-%M-%S", gmtime())
    new_dataset_name = annotation.dataset_name + f"_segmented_{time_str}"
    with wk.webknossos_context("https://webknossos.org"):
        dataset = annotation.get_remote_annotation_dataset()

    volume_mag_view = dataset.layers["Volume"].get_finest_mag()
    mag = volume_mag_view.mag
    volume_annotation_data = volume_mag_view.read(
        absolute_bounding_box=training_data_bbox
    )

    color_mag_view = dataset.layers["color"].mags[mag]

    # Step 2: Initialize a machine learning model to segment our dataset
    features_func = partial(
        feature.multiscale_basic_features, channel_axis=2, edges=False
    )
    segmenter = TrainableSegmenter(features_func=features_func)

    # Step 3: Manipulate our data to fit the ML model and start training on
    # data from our annotated training data bounding box
    print("Starting training…")
    img_data_train = color_mag_view.read(
        absolute_bounding_box=training_data_bbox
    )  # wk data has dimensions (Channels, X, Y, Z)
    # move channels to last dimension, remove z dimension to match skimage's shape
    X_train = np.moveaxis(np.squeeze(img_data_train), 0, -1)
    Y_train = np.squeeze(volume_annotation_data)

    segmenter.fit(X_train, Y_train)

    # Step 4: Use our trained model and predict a class for each pixel in the dataset
    # to get a full segmentation of the data
    print("Starting prediction…")
    X_predict = np.moveaxis(np.squeeze(color_mag_view.read()), 0, -1)
    Y_predicted = segmenter.predict(X_predict)
    segmentation = Y_predicted[:, :, None]  # adds z dimension
    assert segmentation.max() < 256
    segmentation = segmentation.astype("uint8")

    # Step 5: Bundle everything as a WEBKNOSSOS layer and upload to wK for viewing and further work
    with TemporaryDirectory() as tempdir:
        new_dataset = wk.Dataset(
            tempdir, voxel_size=dataset.voxel_size, name=new_dataset_name
        )
        segmentation_layer = new_dataset.add_layer(
            "segmentation",
            wk.SEGMENTATION_CATEGORY,
            segmentation.dtype,
            compressed=True,
            largest_segment_id=int(segmentation.max()),
        )
        segmentation_layer.bounding_box = dataset.layers["color"].bounding_box
        segmentation_layer.add_mag(mag, compress=True).write(segmentation)
        segmentation_layer.downsample(sampling_mode="constant_z")

        remote_ds = new_dataset.upload(
            layers_to_link=[annotation.get_remote_base_dataset().layers["color"]]
            if "PYTEST_CURRENT_TEST" not in os.environ
            else None
        )

    url = remote_ds.url
    print(f"Successfully uploaded {url}")


if __name__ == "__main__":
    main()

This example is inspired by the trainable segmentation example from scikit-image.