Any dataset will require AP, LAT and SEG image paths as input. Corresponding to these images TRANSFORMATIONS has to be defined that convert it to the appropriate Tensor shape.
Currently, this is implemented as follows:
The BaseDataset takes a list of dicts as input data. This may look as follows
paths = [{'ap':'ap.png','lat':'lat.png','seg':'seg.nii.gz'} ]The transforms may be defined as follows:
transforms = {'ap':ap_transform, 'lat':lat_transform, 'seg':seg_transform}Each transform has to be a compose of Dictionary-based MONAI transformations.
lidc_seg_transform = Compose(
[
LoadImageD(
keys={"seg"},
ensure_channel_first=True,
dtype=np.uint8,
simple_keys=True,
image_only=False,
),
SpacingD(
keys={"seg"},
pixdim=(1, 1, 1),
mode="nearest",
padding_mode="zeros",
),
SpatialPadD(keys={"seg"}, spatial_size=(96, 96, 96)),
OrientationD(keys={"seg"}, axcodes="PIR"),
]
)Rib Dataset:
- find views with all rib bones intact (reject cropped view)
- Combine individual rib segments into a single segmentation
- Generate biplanar x-ray to 3D Segmentation dataset
Total Ribs samples: 481
Femur Dataset:
- Collect stats on femur segmentation volume
- Choose all femur segmentation with a volume higher than a threshold. (There are various cropped views of the femur bone. We want to choose those with sufficient bone segments. Crop bones if the segmentation consists of a large section of femur length. We only want the femur head and a few centimetres of the bone.)
- Mirror the right femur so that it looks like the left femur.
Total Femur samples (threshold 50k voxels): 451
Pelvic Dataset:
- Find views with pelvic bones intact (not possible to reject cropped view)
- Generate biplanar x-ray to 3D segmentation dataset
Total Pelvic samples: 446