Ljubljana dataset

PyTorch wrapper for the Ljubljana dataset with parsing of pose parameters

`Ljubljana`

LjubljanaDataset

 LjubljanaDataset (view:str,
                   filename:Union[str,pathlib.Path,NoneType]=None,
                   preprocess:bool=True)

Get X-ray projections and poses from specimens in the Ljubljana dataset.

Given a specimen ID and projection index, returns the projection and the camera matrix for DiffDRR.

	Type	Default	Details
view	str		“ap” or “lat” or “ap-max” or “lat-max”
filename	Union	None	Path to DeepFluoro h5 file
preprocess	bool	True	Preprocess X-rays

Distribution over camera poses

We sample the three rotational and three translational parameters of \(\mathfrak{se}(3)\) from independent normal distributions defined with sufficient variance to capture wide perturbations from the isocenter.

source

get_random_offset

 get_random_offset (view, batch_size:int, device)

Fiducial markers

The LjubljanaDataset class also contains a method for evaluating the registration error for a predicted pose. Digital fiducial markers were placed along the centerlines of the vessels. Projecting them with predicted pose parameters can be used to measure their distance from the true fiducials.

source

Evaluator

 Evaluator (subject, idx)

Initialize self. See help(type(self)) for accurate signature.

Deep learning transforms

We transform X-rays and DRRs before inputting them to a deep learning model by

Rescaling pixels to [0, 1]
Resizing the images to a specified size
Normalizing pixels by the mean and std dev

Code

from tqdm import tqdm

mean, vars = [], []
for view in ["ap", "lat"]:
    specimen = LjubljanaDataset(view, filename="../../data/ljubljana.h5")
    for _, _, _, _, _, _, _, _, _, img, _, _ in tqdm(specimen, ncols=50):
        img = (img - img.min()) / (img.max() - img.min())
        mean.append(img.mean())
        vars.append(img.var())

print("Pixel mean :", sum(mean) / len(mean))
print("Pixel std dev :", (sum(vars) / len(vars)).sqrt())

100%|█████████████| 10/10 [00:12<00:00,  1.25s/it]
100%|█████████████| 10/10 [00:03<00:00,  2.76it/s]

Pixel mean : tensor(0.0774)
Pixel std dev : tensor(0.0569)

source

Transforms

 Transforms (height:int, width:int, eps:float=1e-06)

Transform X-rays and DRRs before inputting to CNN.

Examples

import matplotlib.pyplot as plt
from diffdrr.drr import DRR
from diffdrr.visualization import plot_drr

subject = LjubljanaDataset(view="ap", filename="../../data/ljubljana.h5")

for idx in range(len(subject)):
    (
        volume,
        spacing,
        focal_len,
        height,
        width,
        delx,
        dely,
        x0,
        y0,
        img,
        pose,
        isocenter_pose,
    ) = subject[idx]
    volume[volume < 500] = 0.0
    if idx == 5:
        volume[volume < 1000] = 0.0

    drr = DRR(
        volume,
        spacing,
        focal_len / 2,
        height // 8,
        delx * 8,
        width // 8,
        dely * 8,
        x0,
        y0,
        reverse_x_axis=True,
    ).cuda()
    transforms = Transforms(height // 8, width // 8)

    img = transforms(img).cuda()
    pred_img = drr(None, None, None, pose=pose.cuda())
    pred_img = transforms(pred_img)

    plt.figure(figsize=(12, 3))
    ax = plt.subplot(131)
    plot_drr(img, axs=ax, title="DSA")
    ax = plt.subplot(132)
    plot_drr(pred_img, axs=ax, title="DRR")
    ax = plt.subplot(133)
    plot_drr(img - pred_img, axs=ax, title="Difference", cmap="RdBu")
    plt.show()

subject = LjubljanaDataset(view="lat", filename="../../data/ljubljana.h5")

for idx in range(len(subject)):
    (
        volume,
        spacing,
        focal_len,
        height,
        width,
        delx,
        dely,
        x0,
        y0,
        img,
        pose,
        isocenter_pose,
    ) = subject[idx]
    volume[volume < 500] = 0.0
    if idx == 5:
        volume[volume < 1000] = 0.0

    drr = DRR(
        volume,
        spacing,
        focal_len / 2,
        height // 8,
        delx * 8,
        width // 8,
        dely * 8,
        x0,
        y0,
        reverse_x_axis=True,
    ).cuda()
    transforms = Transforms(height // 8, width // 8)

    img = transforms(img).cuda()
    pred_img = drr(None, None, None, pose=pose.cuda())
    pred_img = transforms(pred_img)

    plt.figure(figsize=(12, 3))
    ax = plt.subplot(131)
    plot_drr(img, axs=ax, title="DSA")
    ax = plt.subplot(132)
    plot_drr(pred_img, axs=ax, title="DRR")
    ax = plt.subplot(133)
    plot_drr(img - pred_img, axs=ax, title="Difference", cmap="RdBu")
    plt.show()