LIDC_IDRI.py

Source: LION/data_loaders/LIDC_IDRI.py

PyTorch dataset adapter for processed LIDC-IDRI axial CT slices.

class LION.data_loaders.LIDC_IDRI.LIDC_IDRI(mode, geometry_parameters=None, parameters=None)

Bases: Dataset

Load LIDC-IDRI data for image-prior, reconstruction, or segmentation tasks.

Parameters:

• mode ({"train", "validation", "test"}) – Dataset split.

• geometry_parameters (Geometry, optional) – Backward-compatible geometry argument.

• parameters (LIONParameter, optional) – Dataset settings returned by default_parameters().

Notes

Processed slices are stored in HU. Image-prior and reconstruction targets are converted to LION normalised intensity. Spatial resizing is performed on CPU before the final device transfer to avoid retaining large 512-square tensors on the GPU for patch training.

static default_parameters(geometry=None, task='reconstruction')

Return default LIDC-IDRI split, sampling, and task parameters.

get_slices_to_load(patient_list, non_nodule_slices_dict, nodule_slices_dict, num_slices_per_patient, pcg_slices_nodule, all_slices_dict=None)

Returns a dictionary that contains patient_id’s as keys and list of slices to load as values for each patient.

Parameters:

• patient_list (List): List that contains patient_id of all patients.

• non_nodule_slices_dict (Dict): Dict that contains all slices without nodule of each patient_id.

• nodule_slices_dict (Dict): Dict that contains all slices with nodule of each patient_id.

• num_slices_per_patient (int): Defines maximum number of slices we want per patient. If num_slices_per_patient=-1 take all slices we have of each patient.

• pcg_slices_nodule (float): Defines amount of slices that should contain a nodule. Value between 0-1.

Returns:

• patient_id_to_slices_to_load_dict which contains patient_id as key and list of slices to load as values

Parameters:

• patient_list (List)

• non_nodule_slices_dict (Dict)

• nodule_slices_dict (Dict)

• num_slices_per_patient (int)

• pcg_slices_nodule (float)

• all_slices_dict (Dict | None)

get_patient_id_to_first_index_dict(patient_with_slices_to_load)

Returns a dictionary that contains patient_id’s as keys and start index of each patient in self.slice_index_to_patient_id_list as value.

Parameters:

• patient_with_slices_to_load (Dict): Dict that defines which slices to load per patient.

Returns:

• patient_id_to_first_index_dict (Dict): Defines start index of each patient in self.slice_index_to_patient_id_list. Needed for mapping of global index to slice index.

Parameters:

patient_with_slices_to_load (Dict)

get_slice_index_to_patient_id_list(patient_with_slices_to_load)

Returns a list that contains “number of slices” times each patient id.

Parameters:

• patient_with_slices_to_load (Dict): Dict that defines which slices to load per patient.

Returns:

• slice_index_to_patient_id_list (List): Contains number of slices times each patient id. Needed for mapping of global index to slice index.

Parameters:

patient_with_slices_to_load (Dict)

get_reconstruction_tensor(file_path)

Load and resize one processed HU slice for reconstruction use.

Parameters:

file_path (Path)

Return type:

Tensor

set_sinogram_transform(sinogram_transform)

Set an optional transform applied to generated sinograms.

set_image_transform(image_transform)

Set an optional transform applied to image targets.

compute_clean_sinogram(image=None)

Forward-project an image with the dataset CT operator.

Return type:

Tensor

get_mask_tensor(patient_id, slice_index)

Load the random or consensus nodule mask for one slice.

Parameters:

• patient_id (str)

• slice_index (int)

Return type:

Tensor