Experimental Design

Experimental Design

Ideally, the purpose of the algorithm is to aid a medical professional in calculating the volume of a pulmonary nodule. While examining the CT scan of an image, the user may select a portion of the image in which the lesion is the largest object. To simulate this, CT images from the course database served as the test data. The VisionX tool vclip was used to separate the portion of the image containing the lesion. Other test inputs included synthetically generated byte images as well as synthetic images with added Gaussian noise. Various VisionX programs were utilized to create this test images, most notably vgenim and vgen3d.

The test images are as follows:

  • test_sphere.vx - a spherical object, as well as gridlines to simulate blood vessels surrounding the lung nodule
  • test_cube.vx - a translated, rotated cube
  • test_hollow.vx - a cube with a spherical hollow inside
  • test_impulse.vx - a standard impulse test function, consisting of a single point
  • test_ellipse.vx - a rotated ellipse

Full Run-Through of Volume Calculation of Pulmonary Nodule

Lung CT scan cross section with highlighted nodule

test_cube.vx 3-dimensional test image

test_ellipse.vx 3-dimensional test image

Additionally, Gaussian noise was added to each image using the vpix program to double the number of synthetic test images. For each test image, the initial resolution (voxel size) is noted. Synthetic test images are assigned a voxel size of .7x.7x1.25 as this closely represents typical voxel sizes seen in CT images. The images are then interpolated to a smaller voxel size (.3x.3x.4 was chosen) using the interpolation methods of choice. Segmentation is then performed on all images. This process involves iterative thresholding followed by morphological filtering for noise removal and vessel cutting as outlined here.

After running the program on test images with and without noise, the process was applied to actual CT scan images containing pulmonary nodules. Five images were selected as good candidates to evaluate the interpolators due to the high visibility of the nodule as well as lack of obstruction. They were obtained from the ECE 5470 course database:

    • ct1 - /classes/ece547/lab/lab6/ctimage.vs
    • ct2 - /course-data/apr5/CP/SC0021/bSC0001-153013-04-01.vs
    • ct3 - /course-data/apr5/CP/SC0016/bSC0001-171645-02-01.vs
    • ct4 - /course-data/apr5/CP/SC0019/bSC0001-091746-02-05.vs
    • ct5 - /course-data/apr5/CP/SC0016/bSC0001-180106-02-01.vs

We measure the ”quality” of the interpolator by its affect on the accuracy of the volume calculation. For synthetic images, the actual volume is exactly known from the command-line arguments. For the CT images, we were careful to select test data that contained boundary markings from previous users in the study, using these .vxa files to define the ”correct” segmentation for volume calculation. The figures on the right detail the steps taken in performing a full volume calculation on a given CT image.

For a proper understanding of how the experimental process is carried out, inspection of the program code and scripts used is necessary. All the project code and documentation can be found here.

Interpolated user selected nodule region

Iteratively thresholded nodule region

Morphologically filtered nodule region

3D representation of nodule region