data

Load DICOM datasets as numpy arrays with voxel dimensions

CT scans in DiffDRR are stored using the torchio.Subject dataclass. torchio provides a convenient and consistent mechanism for reading volumes from a variety of formats and orientations. We canonicalize all volumes to the RAS+ coordinate space. In addition to reading an input volume, you can also pass the following to diffdrr.data.read when loading a subject:

labelmap : a 3D segmentation of the input volume
labels : a subset of structures from the labelmap that you want to render
orientation : a frame-of-reference change for the C-arm (currently, “AP” and “PA” are supported)
bone_attenuation_multiplier : a constant multiplier to the estimated density of bone voxels
fiducials : a tensor of 3D fiducial marks in world coordinates
**kwargs : any additional kwargs can be passed to the torchio.Subject and accessed as a dictionary

Density conversion and the thorax

Voxels are thresholded at -800 and 350 HU into air, soft tissue, and bone, and every voxel at or below -800 HU is assigned a single value before the volume is min-max normalized. This convention is inherited from DeepDRR and is appropriate where bone carries the signal.

On thoracic CT it is not. Normal aerated lung parenchyma spans roughly -950 to -700 HU, so most of the lung is mapped onto the floor and becomes indistinguishable from the air outside the patient; low-attenuation findings such as emphysema are lost entirely. If you need lung attenuation preserved, for instance when rendering training data for chest reconstruction, use nanodrr, whose hu_to_mu clamps at -1000 HU and applies a bilinear water/bone model (#378).

source

load_example_ct

def load_example_ct(
    labels:NoneType=None, orientation:str='AP', bone_attenuation_multiplier:float=1.0, **kwargs
)->Subject:

Load an example chest CT for demonstration purposes.

source

read

def read(
    volume:str | Path | ScalarImage, # CT volume
    labelmap:str | Path | LabelMap=None, # Labelmap for the CT volume
    labels:int | list=None, # Labels from the mask of structures to render
    orientation:str | None='AP', # Frame-of-reference change
    bone_attenuation_multiplier:float=1.0, # Scalar multiplier on density of high attenuation voxels
    fiducials:torch.Tensor=None, # 3D fiducials in world coordinates
    transform:RigidTransform=None, # RigidTransform to apply to the volume's affine
    center_volume:bool=True, # Move the volume's isocenter to the world origin
    resample_target:NoneType=None, # Resampling resolution argument passed to torchio.transforms.Resample
    **kwargs
)->Subject: # Any additional information to be stored in the torchio.Subject

Read an image volume from a variety of formats, and optionally, any given labelmap for the volume. Converts volume to a RAS+ coordinate system and moves the volume isocenter to the world origin.