Radonc/BME Medical Image and
Computational Analysis Laboratory
Sam Martocci
MS candidate in BME
expected graduation Spring 2022


email link


Research Overview

Effect of Pixel Intensity Threshold and Imaging Parameters on Lung Vessel Segmentation from Chest CT

Background and Significance
Lung vessel quantification possesses significant implications for the detection, diagnosis and further treatment of many vascular diseases including chronic pulmonary arterial hypertension, chemotherapy-related lung or tumor injury, altered vascular development in children with extreme pre-term gestation, injury due to radiation or other environmental insult, and angiogenesis around tumors. Accuracy in identification and characterization of lung vessels from chest CT is dependent on parameters such as voxel size, image reconstruction parameters, patient motion, and use of a vascular contrast.

Methods
A representative chest CT scan of a human subject was acquired and the raw data, previously saved on the CT workstation, was used to reconstruct and generate a set of 16 CT data-sets: slice thicknesses: 0.5, 1, 2 and 3 mm (no overlap); in-plane pixel dimension: 0.543 and 0.702 mm; and with/without a Lung enhancement filter. For each of the image sets, the 3D lung vascular system was segmented and characterized semi-automatically via flood-filling starting from a manually-selected seed point in the pulmonary root and a pixel-intensity threshold for distinguishing vessel from background1,2. The same seed point and thresholds were used for all runs. A radius-histogram for vessel count was tabulated for each data set. A calibration model was formulated as: N = No (1 + a1S + a2P + a3F+ a4S2 + a5SF) where S = slice thickness[mm], P = in-plane pixel size[mm] and F = {0,1} for Body vs. Lung filter.

Results
Overall, there is a trend towards increased blood vessels detected for lower thresholds. This makes sense because as the pixel intensity requirements are lowered, we expect that more pixels will be accepted and characterized as lung vessels. This trend was observed much more frequently at lower radii. At slice thickness of 0.5, there was a difference of approximately 9.4% in the vessel count with a 20 Hu difference in threshold (data not shown). Since the scans with lower slice thickness can detect smaller vessels, it is expected that the threshold intensity will have a larger effect on these scans.
Figure 1. Plot illustrating number of detected vessel branches when using the -560 vs. -580 threshold on the Hounsfield scale for CT scans. In this graph the -580 threshold picked up approximately 300 more vessels at the lowest radius than at the -560 threshold. This graph shows the effect of threshold for a slice thickness of 1 mm, in-plane pixel dimension of 0.543 mm, and with the lung filter applied.


Conclusions
Vessel branch counts have large differences with varying pixel intensity threshold and different scan parameters. Hence, these differences in voxel size, slice thickness, reconstruction filters and thresholds for pixel intensities need to be analyzed to create a way for quantitative comparison of metrics across scans. More research is needed to mathematically model the effects of increasing the threshold difference.


  • National Presentations
    Martocci S, Siva Kumar S, O' Dell W.
          Effect of Pixel Intensity Threshold and Imaging Parameters on Lung Vessel Segmentation from Chest CT.
          Biomedical Engineering Society Annual Meeting. (Orlando, Fl):
          Oct. 7, 2021.

    Martocci S, Siva Kumar S, O' Dell W.
          Effect of Imaging Parameters on Lung Vessel Segmentation from Chest CT Scans.
          Biomedical Engineering Society Annual Meeting. (virtual):
          Oct. 16, 2020.

  • Related MIACALab projects
    1. Lung radiation dose response

  • Close Research Collaborators
    1. W. O'Dell (Radiation Oncology Research)
    2. Dr. Julie Bradley
    3. Dr. Paul Okunieff