linumpy.gpu.registration#

GPU-accelerated registration operations for linumpy.

Provides a hybrid approach where metric computation is done on GPU while the optimizer runs on CPU (SimpleITK).

Classes#

GPUAcceleratedRegistration

Hybrid GPU/CPU registration class.

Functions#

register_2d_gpu(fixed, moving[, method, metric, ...])

GPU-accelerated 2D image registration.

apply_transform_gpu(image, transform[, use_gpu])

Apply SimpleITK transform to image using GPU resampling.

Module Contents#

class linumpy.gpu.registration.GPUAcceleratedRegistration(use_gpu=True, metric='mse')[source]#

Hybrid GPU/CPU registration class.

Uses GPU for: - Image resampling/transformation - Metric computation (MSE, NCC)

Uses CPU (SimpleITK) for: - Optimization loop - Transform management

Parameters:

  • use_gpu (bool) – Whether to use GPU for metric computation

  • metric (str) – Registration metric: ‘mse’, ‘ncc’, ‘mi’

use_gpu = False[source]#
metric = ''[source]#
compute_metric(fixed, moving)[source]#

Compute registration metric between two images.

Parameters:

  • fixed (np.ndarray) – Fixed image

  • moving (np.ndarray) – Moving image (already transformed)

Returns:

Metric value (lower is better for MSE, higher for NCC)

Return type:

float

transform_image(image, transform_matrix, output_shape=None)[source]#

Apply transformation to image using GPU.

Parameters:

  • image (np.ndarray) – Input image

  • transform_matrix (np.ndarray) – Transformation matrix

  • output_shape (tuple, optional) – Output shape

Returns:

Transformed image

Return type:

np.ndarray

linumpy.gpu.registration.register_2d_gpu(fixed, moving, method='affine', metric='mse', max_iterations=1000, use_gpu=True)[source]#

GPU-accelerated 2D image registration.

Uses SimpleITK optimizer with GPU metric computation.

Parameters:

  • fixed (np.ndarray) – Fixed image

  • moving (np.ndarray) – Moving image

  • method (str) – Transform type: ‘translation’, ‘euler’, ‘affine’

  • metric (str) – Metric: ‘mse’, ‘ncc’, ‘mi’

  • max_iterations (int) – Maximum optimizer iterations

  • use_gpu (bool) – Whether to use GPU acceleration

Returns:

  • transform (sitk.Transform) – Computed transform

  • str – Optimizer stop condition

  • float – Final metric value

Return type:

Any

linumpy.gpu.registration.apply_transform_gpu(image, transform, use_gpu=True)[source]#

Apply SimpleITK transform to image using GPU resampling.

Parameters:

  • image (np.ndarray) – Input image

  • transform (sitk.Transform) – SimpleITK transform

  • use_gpu (bool) – Whether to use GPU

Returns:

Transformed image

Return type:

np.ndarray