miblab.kidney_pc_dixon#

miblab.kidney_pc_dixon(input_array, model='unetr', overlap=0.3, postproc=True, clear_cache=False, verbose=False)[source]#

Segment individual kidneys on post-contrast Dixon images.

This requires 4-channel input data with out-phase images, in-phase images, water maps, and fat maps.

Parameters:

  • input_array (numpy.ndarray) – A 4D numpy array of shape [x, y, z, contrast] representing the input medical image volume. The last index must contain out-phase, in-phase, water and fat images, in that order.

  • model –

    model = ‘unetr’ uses a pretrained UNETR-based model in MONAI, hosted on Zenodo model = ‘nnunet’ uses a pretrained nnUNet-based model, hosted on Zenodo under the hood this runs nnUNetPredictor (for more details MIC-DKFZ Wiki)

  • overlap (float) – only valid for model = ‘unetr’ defines the amount of overlap between adjacent sliding window patches during inference. A higher value (e.g., 0.5) improves prediction smoothness at patch borders but increases computation time.

  • postproc (bool) – If True, applies post-processing to select the largest connected component from the UNETR output for each kidney mask

  • clear_cache – If True, the downloaded pth file is removed again after running the inference.

  • verbose (bool) – If True, prints logging messages.

Returns:

A dictionary with the keys ‘leftkidney’ and ‘rightkidney’, each containing a binary NumPy array representing the respective kidney mask.

Return type:

dict

Example

>>> import numpy as np
>>> import miblab
>>> data = np.random.rand(128, 128, 30, 4)
>>> mask = miblab.kidney_pc_dixon(data)
>>> print(mask['kidney_left'])
[0 1 1 ... 0 0 0]
On this page

This Page