Daily QA CBCT Image Processing

This describes the BB finding algorithm and precise measurement of the BB's location.

Coarse (approximate) Location of BB

The first step is to determine the BB's approximate location within the CBCT volume. The method needs to be immune to image artifacts and to avoid voxels occupied by the couch. To do this, the algorithm first locates the phantom cube and then the BB within the cube.

A 3D volume is constructed using the CBCT slices. Next, the volume is divided into horizontal slices parallel to the table and perpendicular to the Y-axis. Starting at the top, slices are examined in descending order to determine if they contain the cube.

The examination of a single slice involves the following steps:

• Construct the slice from the 3D volume. The slice is parallel to the table/couch and is orthogonal to the originally acquired slices.
• Remove noise. This is done by creating a histogram of the values (HU levels). Images containing the cube have been observed to have 600 different voxel values. Values that fall in the lower portion of the histogram are deemed noise, and are set to zero. The size of the portion of the histogram is driven by the configured value DailyQACBCTDarkPixelValueLevels .
  
  

  Images that are mostly 'black' will be made completely black. Images that contain the cube or the table will have the dark parts of the images made completely black.

• Noise is reduced further by finding the sum of the count in histogram of the brightest voxels that is sufficient to account for the area of the cube (top view). Values in the histogram below this value are also considered noise and all voxels in the image with these low values are set to zero.
• The slice is then examined to find rising and falling edges that are separated by the cube's approximate size. Edge pairs are found one row or column of voxels at a time. A rising edge is defined as transitioning from zero to non-zero, and a falling edge is the reverse. If the approximate correct number of pairs are found in both orthogonal directions (X and Z in RTPLAN space), then the slice contains the cube.

  Note that using transitions is necessary because the profile of the cube is occasionally not symmetrical. This asymmetry thwarts approaches such as center of mass and edge finding using the severity of slope.

When two adjacent slices are found where the higher one does not contain the cube and the lower one does, then the top of the cube has been established. The vertical center of the cube is determined by descending a further 1/2 the cube's height. Several slices are checked at the vertical midpoint to increase confidence that the cube has been found.

The horizontal center of the cube is found using the midpoint between edges found in the Z and X directions. At this point the coarse (approximate) position of the cube has been found, which is accurate to within 2 voxels.

Fine (precise) Location of BB

The following steps are taken to determine the precise location of the BB.

• A sub-volume is taken that surrounds the coarse location.
• The sums of the planes in each axis of the sub-volume are summed, producing three one-dimensional arrays.
• For each one-dimensional array, a cubic spline is calculated.
• The maximum point in the cubic spline is found, which corresponds to the center of the BB in that dimension.
• To qualify as the BB, the standard deviation of the maximum must be above a threshold (specified in configuration file). This is necessary because any profile, even without the BB, will have a maximum due to noise, and such cases must be rejected. For this reason, taking a large enough volume around the BB handles two problems:
  • Accommodate small errors from the coarse find.
  • Provide a large enough sample of voxels so as to establish a meaningful standard deviation.
  The volume can not be too large though, or its outer edges will be outside the cube and corrupt the standard deviation, making it meaningless.