zea.data.augmentations

Augmentation layers for ultrasound data.

Classes

RandomCircleInclusion(radius[, fill_value, ...])

Adds a circular inclusion to the image, optionally at random locations.
class zea.data.augmentations.RandomCircleInclusion(radius, fill_value=1.0, circle_axes=(1, 2), with_batch_dim=True, return_centers=False, recovery_threshold=0.1, randomize_location_across_batch=True, seed=None, width_range=None, height_range=None, **kwargs)[source]

Bases: Layer

Adds a circular inclusion to the image, optionally at random locations.

Since this can accept N-dimensional inputs, you’ll need to specify your circle_axes – these are the axes onto which a circle will be drawn. This circle will then be broadcast along the remaining dimensions.

You can then optionally specify whether there is a batch dim, and whether the circles should be located randomly across that batch.

For example, if you have a batch of videos, e.g. of shape [batch, frame, height, width], then you might want to specify circle_axes=(2, 3), and randomize_location_across_batch=True. This would result in a circle that is located in the same place per video, but different locations for different videos.

Once your method has recovered the circles, you can evaluate them using the evaluate_recovered_circle_accuracy() method, which will expect an input shape matching your inputs to call().

Initialize RandomCircleInclusion.

Parameters:

  • radius (int | tuple[int, int]) – Radius of the circle/ellipse to include.

  • fill_value (float) – Value to fill inside the circle.

  • circle_axes (tuple[int, int]) – Axes along which to draw the circle (height, width).

  • with_batch_dim (bool) – Whether input has a batch dimension.

  • return_centers (bool) – Whether to return circle centers along with images.

  • recovery_threshold (float) – Threshold for considering a pixel as recovered.

  • randomize_location_across_batch (bool) – If True (and with_batch_dim=True), each batch element gets a different random center. If False, all batch elements share the same center.

  • seed (Any) – Optional random seed for reproducibility.

  • width_range (tuple[int, int]) – Range (min, max) for circle center x (width axis).

  • height_range (tuple[int, int]) – Range (min, max) for circle center y (height axis).

  • **kwargs – Additional keyword arguments for the parent Layer.

Example

>>> from zea.data.augmentations import RandomCircleInclusion
>>> from keras import ops

>>> layer = RandomCircleInclusion(
...     radius=5,
...     circle_axes=(1, 2),
...     with_batch_dim=True,
... )
>>> image = ops.zeros((1, 28, 28), dtype="float32")
>>> out = layer(image)
build(input_shape)[source]

Build the layer and compute static shape and permutation info.

Parameters:

input_shape (tuple) – Shape of the input tensor.

call(x, seed=None)[source]

Apply the random circle inclusion augmentation.

Parameters:

  • x (Tensor) – Input tensor.

  • seed (Any, optional) – Optional random seed for reproducibility.

Returns:

Augmented images, and optionally the circle

centers if return_centers is True.

Return type:

Tensor or tuple

compute_output_shape(input_shape)[source]

Compute output shape for the layer.

Parameters:

input_shape (tuple) – Shape of the input tensor.

Returns:

The output shape (same as input).

Return type:

tuple

evaluate_recovered_circle_accuracy(images, centers, recovery_threshold, fill_value=None)[source]

Evaluate the percentage of the true circle that has been recovered in the images, and return a mask of the detected part of the circle.

Parameters:

  • images (Tensor) – Tensor of images (any shape, with circle axes as specified).

  • centers (Tensor) – Tensor of circle centers (matching batch size).

  • recovery_threshold (float) – Threshold for considering a pixel as recovered.

  • fill_value (float, optional) – Optionally override fill_value for cases where image range has changed.

get_config()[source]

Get layer configuration for serialization.

Returns:

Dictionary of layer configuration.

Return type:

dict