fbp#

diffct.reconstruction.fbp(sinogram, trajectory, volume_shape, angles, *, beam='parallel', detector_spacing=1.0, voxel_spacing=1.0, window='ram-lak', pad_factor=2, projection_weights=None, angular_weights=None, filter_response=None, short_scan=False, isocenter=None, backend='auto', precision='fast', devices=None, view_chunk_size=None, volume_chunk_shape=None, distributed=False, process_group=None, partition='views', partition_grid=None)[source]#

Differentiable parallel/fan FBP for fixed flat-detector geometry.

sinogram is a floating tensor (views, bins); volume_shape is (H,W). angles are explicit unwrapped radians. Detector/voxel spacings share the trajectory’s physical unit. projection_weights and angular_weights replace the default cosine and angular quadrature, respectively. filter_response replaces the complete real rFFT multiplier, including its physical units; length is pad_factor*bins//2+1. Its use ignores window. These tensors retain autograd. short_scan=True adds Parker weighting before filtering for a circular fan scan with sufficient coverage. precision="fast" preserves native FP32 execution; "float32" and "float64" select PyTorch arithmetic in that dtype. Output is signed and stays on the input device. View/spatial chunks, local devices and views/space/hybrid rank ownership are supported. Provide full global geometry/angles and the owned view data in distributed mode. Geometry gradients are rejected. Use fbp_into() for numerical block stores.