dosma.models package

Submodules

dosma.models.oaiunet2d module

@author: Arjun Desai

  1. Stanford University, 2019

class dosma.models.oaiunet2d.IWOAIOAIUnet2D(input_shape, weights_path, force_weights=False)[source]

Bases: dosma.models.oaiunet2d.OAIUnet2D

Model trained by Team 6 in the 2019 IWOAI Segmentation Challenge.

References

Desai, et al., “The International Workshop on Osteoarthritis Imaging Knee MRI Segmentation Challenge: A Multi-Institute Evaluation and Analysis Framework on a Standardized Dataset.” arXiv preprint arXiv:2004.14003 (2020). [link]

ALIASES = ['iwoai-2019-t6']
generate_mask(volume: dosma.data_io.med_volume.MedicalVolume)[source]

Segment the MRI volumes.

Parameters

volume (MedicalVolume) – Volume to segment in proper orientation.

Returns

Volumes are binarized (0,1)

uint8 3D ndarray of shape volume.shape.

Return type

MedicalVolume or List[MedicalVolume]

Raises

ValueError – If volumes is not 3D ndarray or if tissue is not a string or not in list permitted tissues.

class dosma.models.oaiunet2d.IWOAIOAIUnet2DNormalized(input_shape, weights_path, force_weights=False)[source]

Bases: dosma.models.oaiunet2d.IWOAIOAIUnet2D

Extension of model trained by Team 6 in the 2019 IWOAI Segmentation Challenge (with normalization).

This model uses the same architecture as IWOAIOAIUnet2D, but pre-processes the input data by zero-mean, unit-std normalization.

References

Desai, et al., “The International Workshop on Osteoarthritis Imaging Knee MRI Segmentation Challenge: A Multi-Institute Evaluation and Analysis Framework on a Standardized Dataset.” arXiv preprint arXiv:2004.14003 (2020).

ALIASES = ('iwoai-2019-t6-normalized',)
class dosma.models.oaiunet2d.OAIUnet2D(input_shape, weights_path, force_weights=False)[source]

Bases: dosma.models.seg_model.KerasSegModel

Model trained in Chaudhari et al. IWOAI 2018

Original Github: https://github.com/akshaysc/msk_segmentation

ALIASES = ['oai-unet2d', 'oai_unet2d']
generate_mask(volume: dosma.data_io.med_volume.MedicalVolume)[source]

Segment the MRI volumes.

Parameters

volume (MedicalVolume) – Volume to segment in proper orientation.

Returns

Volumes are binarized (0,1)

uint8 3D ndarray of shape volume.shape.

Return type

MedicalVolume or List[MedicalVolume]

Raises

ValueError – If volumes is not 3D ndarray or if tissue is not a string or not in list permitted tissues.

sigmoid_threshold = 0.5

dosma.models.seg_model module

Abstract classes to wrap Keras model.

T SegModel: Abstract wrapper for Keras model used for semantic segmentation.

class dosma.models.seg_model.KerasSegModel(input_shape, weights_path, force_weights=False)[source]

Bases: dosma.models.seg_model.SegModel

Abstract wrapper for Keras model used for semantic segmentation

build_model(input_shape, weights_path)[source]

Builds a segmentation model architecture and loads weights.

Parameters

  • input_shape – Input shape of volume

  • weights_path –

Returns

a segmentation model that can be used for segmenting tissues (a Keras/TF/PyTorch model)

class dosma.models.seg_model.SegModel(input_shape, weights_path, force_weights=False)[source]

Bases: abc.ABC

Parameters

  • input_shape (Tuple[int] or List[Tuple[int]]) – Shapes(s) for initializing input(s) into model in format (height, width, channels).

  • weights_path (str) – filepath to weights used to initialize Keras model

  • force_weights (bool, optional) – If True, load weights without any checking. Keras/Tensorflow only.

ALIASES = ['']
abstract build_model(input_shape, weights_path)[source]

Builds a segmentation model architecture and loads weights.

Parameters

  • input_shape – Input shape of volume

  • weights_path –

Returns

a segmentation model that can be used for segmenting tissues (a Keras/TF/PyTorch model)

abstract generate_mask(volume: dosma.data_io.med_volume.MedicalVolume)[source]

Segment the MRI volumes.

Parameters

volume (MedicalVolume) – Volume to segment in proper orientation.

Returns

Volumes are binarized (0,1)

uint8 3D ndarray of shape volume.shape.

Return type

MedicalVolume or List[MedicalVolume]

Raises

ValueError – If volumes is not 3D ndarray or if tissue is not a string or not in list permitted tissues.

dosma.models.seg_model.whiten_volume(x: numpy.ndarray, eps: float = 0.0)[source]

Whiten volumes by mean and std of all pixels.

Parameters

  • x (ndarray) – 3D numpy array (MRI volumes)

  • eps (float, optional) – Epsilon to avoid division by 0.

Returns

A numpy array with mean ~ 0 and standard deviation ~ 1

Return type

ndarray

dosma.models.util module

Functions for loading Keras models

@author: Arjun Desai

  1. Stanford University, 2019

dosma.models.util.get_model(model_str, input_shape, weights_path, **kwargs)[source]

Get a Keras model :param model_str: model identifier :param input_shape: tuple or list of tuples for initializing input(s) into Keras model :param weights_path: filepath to weights used to initialize Keras model :return: a Keras model

Module contents

class dosma.models.IWOAIOAIUnet2D(input_shape, weights_path, force_weights=False)[source]

Bases: dosma.models.oaiunet2d.OAIUnet2D

Model trained by Team 6 in the 2019 IWOAI Segmentation Challenge.

References

Desai, et al., “The International Workshop on Osteoarthritis Imaging Knee MRI Segmentation Challenge: A Multi-Institute Evaluation and Analysis Framework on a Standardized Dataset.” arXiv preprint arXiv:2004.14003 (2020). [link]

ALIASES = ['iwoai-2019-t6']
generate_mask(volume: dosma.data_io.med_volume.MedicalVolume)[source]

Segment the MRI volumes.

Parameters

volume (MedicalVolume) – Volume to segment in proper orientation.

Returns

Volumes are binarized (0,1)

uint8 3D ndarray of shape volume.shape.

Return type

MedicalVolume or List[MedicalVolume]

Raises

ValueError – If volumes is not 3D ndarray or if tissue is not a string or not in list permitted tissues.

class dosma.models.IWOAIOAIUnet2DNormalized(input_shape, weights_path, force_weights=False)[source]

Bases: dosma.models.oaiunet2d.IWOAIOAIUnet2D

Extension of model trained by Team 6 in the 2019 IWOAI Segmentation Challenge (with normalization).

This model uses the same architecture as IWOAIOAIUnet2D, but pre-processes the input data by zero-mean, unit-std normalization.

References

Desai, et al., “The International Workshop on Osteoarthritis Imaging Knee MRI Segmentation Challenge: A Multi-Institute Evaluation and Analysis Framework on a Standardized Dataset.” arXiv preprint arXiv:2004.14003 (2020).

ALIASES = ('iwoai-2019-t6-normalized',)
class dosma.models.OAIUnet2D(input_shape, weights_path, force_weights=False)[source]

Bases: dosma.models.seg_model.KerasSegModel

Model trained in Chaudhari et al. IWOAI 2018

Original Github: https://github.com/akshaysc/msk_segmentation

ALIASES = ['oai-unet2d', 'oai_unet2d']
generate_mask(volume: dosma.data_io.med_volume.MedicalVolume)[source]

Segment the MRI volumes.

Parameters

volume (MedicalVolume) – Volume to segment in proper orientation.

Returns

Volumes are binarized (0,1)

uint8 3D ndarray of shape volume.shape.

Return type

MedicalVolume or List[MedicalVolume]

Raises

ValueError – If volumes is not 3D ndarray or if tissue is not a string or not in list permitted tissues.

sigmoid_threshold = 0.5
dosma.models.get_model(model_str, input_shape, weights_path, **kwargs)[source]

Get a Keras model :param model_str: model identifier :param input_shape: tuple or list of tuples for initializing input(s) into Keras model :param weights_path: filepath to weights used to initialize Keras model :return: a Keras model