Radonc/BME     Medical Image and Computational Analysis Laboratory
Segmenting and characterizing pulmonary vasculature tree structure. Investigators:
W. O'Dell, PhD, Dept. of Radiation Oncology and Biomedical Engineering
Drs. Ali Ataya, Hassan Alnuaimat and Jorge Lascano, UF Dept. of Pulmonology
Sreekala Prabhakaran, MD, UF Dept. of Pediatrics, Pediatric Pulmonary Division
R. Jim White, MD, PhD, Dept. of Pulmonology, University of Rochester
We were initially interested separating the pulmonary vascular structures from the regions of radiation fibrosis in the lung for radiation dose response analysis (see that project page here). Dr. R. Jim White at URMC approached us after Sindhuja TG's MS thesis presentation to apply this approach to automatically quantify changes in pulmonary arterial vasculature in a rat model of chronic pulmonary arterial hypertension (PAH). During Ankit Salgia's MS thesis presentation we were asked to extended this work to analysis of pulmonary vascular development in human neonated with extreme pre-term gestation. We are currently applying non-invasive assessment of pulmonary vascular structure to quantify (1) the response of the lung to radiation exposure in breast cancer patients (funded through a grant from the Florida Department of Health), (2) the development of lung vasculature in children born extremely prematurely (student support [
Aren Saini and Alex Ozycz] via the UF Research Scholars Program), and (3) changes in vascular anatomy adults and in rats lungs with various vascular diseases.
A healthy adult human contrast-enhanced chest CT data set. [A] Image A is a maximum intensity projection (MIP) of the CT dataset of the right hemi-lung after extraction of the lung volume. This is a side view of the chest looking outward from the perspective of the heart with the lung apex at the top. [B] Image B is a depth-enhanced MIP of the extracted vessel tree structure resulting from a region-growing algorithm seeded at the root of one of the large vessels in the lower right hemi-lung. The distinct vascular beds of the 3 right lung lobes are apparent in this view. [C] Image C is a depth-enhanced MIP of the skeletonized tree in the same orientation. [D] Image D is a colorized MIP of the simulated image of the 8 largest vessel trees resulting from the splitting of each connected tree. The corrected tree structure was found to have 1,960 branches on 32 independent trees.
[A] 3D pulmonary arterial tree segmentation result for a healthy rat hemi-lung. Micro-CT images were obtained at approximately 40 micron isotropic resolution.
[B] 3D pulmonary arterial tree segmentation result for a rat with experimentally-induced PAH and demonstrating marked changes/remodeling of the pulmonary arterial tree.
[C] Log-histogram of number of branches versus branch radius for the healthy and PHA rat.
We discovered that threshold-based methods of vessel segmentation lead to underestimation of vessel radii. We then adapted (and patented) a method for automatic tumor sizing to vessel sizing. This approach was recently validated using a realistic 3D physical phantom. The phantom was created by taking the segmented vascular tree from Figure 1C, and creating a mathematical representation of the largest tree wherein each branch centerline is repsentated as a spline-based curve in 3D space and the branch surface as a constant-radius tube structure. Branches were connected using a mathematical model of the bifurcation region. From this mathematical model we generated a 3D surface mesh of the tree which was outputted as a stereolithography file what was send to a 3D printer. We then took the 3D printout of the human lung vascular tree and manually measured 69 branches. This tree was then put into a commercial clinical CT scanner and we processed those images using our software, to compare against the manual measurements. This work was shown by student Anne Gormaley at the national BMES conference in 2017 and was published in the journal Medical Physics.