CT-derived vessel segmentation for analysis of post-radiation therapy changes in vasculature and perfusion

Abstract

Vessel segmentation in the lung is an ongoing challenge. While many methods have been able to successfully identify vessels in normal, healthy, lungs, these methods struggle in the presence of abnormalities. Following radiotherapy, these methods tend to identify regions of radiographic change due to post-radiation therapytoxicities as vasculature falsely. By combining texture analysis and existing vasculature and masking techniques, we have developed a novel vasculature segmentation workflow that improves specificity in irradiated lung while preserving the sensitivity of detection in the rest of the lung. Furthermore, radiation dose has been shown to cause vascular injury as well as reduce pulmonary function post-RT. This work shows the improvements our novel vascular segmentation method provides relative to existing methods. Additionally, we use this workflow to show a dose dependent radiation-induced change in vasculature which is correlated with previously measured perfusion changes (R ² = 0.72) in both directly irradiated and indirectly damaged regions of perfusion. These results present an opportunity to extend non-contrast CT-derived models of functional change following radiation therapy.

Keywords: ct-derived perfusion; lung perfusion; post-RT vascular change; pulmonary vasculature segmentation; radiation-induced damage.

FIGURE 1

Delivered radiation dose distributions for WMS groups (A) and (B) show treatment differences. Group B had a more medial and cranial dose distribution with a smaller region of targeted high dose values.

FIGURE 2

Example of a post-RT CT scan showing radiographic damage (left frame), the resulting vessel segmentation (middle frame), and the texture classification for each voxel (right frame) that was classified as a vessel. The vessel segmentation classifies several voxels in the area of the CT showing radiographic change as vessel that are likely false positives. Texture analysis was performed over the entire lung and is shown masked by the voxels that were identified as vessel in the third panel. The textures of the voxels in the region of false-positives are bronchovascular, consolidated, or ground-glass reticular which are expected radiation-induced textures.

FIGURE 3

Flowchart showing the workflow to produce vascular segmentations using the tools described.

FIGURE 4

Flowchart showing the workflow to perform the indirect vascular change analysis using the tools described.

FIGURE 5

Example of the vessel segmentation algorithm results for a pre-RT (top row) and post-RT (bottom row) example. From left to right: the original CT, the vessel segmentation denoting voxels classified as vessels in green, the result of removing voxels that were classified as both vessels and either ground glass reticular or bronchovascular (red), and finally the result of adding back in the large vessels (blue).

FIGURE 6

Segmentation results in the five Group (A) swine. Each column represents a subject where the top row shows the post-RT CT image in an axial slice showing the post-RT radiographic change. The middle row shows the original vessel segmentation overlayed on the CT in red which in all subjects classified damaged regions of the lung as vessel. The bottom row shows the result of the novel vessel segmentation workflow overlayed on the CT in green. In all subjects, the apparent quality of the vessel segmentation improves in the regions of radiographic change as indicated by the reduction in large connected regions being identified. The resulting segmentation appears to align with vessels that can be observed in the CT in both irradiated and non-irradiated regions.

FIGURE 7

Segmentation results in the five Group (B) swine. Each column represents a subject where the top row shows the post-RT CT image in an axial slice showing the post-RT radiographic change. The middle row shows the original vessel segmentation overlayed on the CT in red which in all subjects classified damaged regions of the lung as vessel. The bottom row shows the result of the novel vessel segmentation workflow overlayed on the CT in green. In all subjects, the apparent quality of the vessel segmentation improves in the regions of radiographic change as indicated by the reduction in large connected regions being identified. The resulting segmentation appears to align with vessels that can be observed in the CT in both irradiated and non-irradiated regions.

FIGURE 8

Example cropped 3D Rendering of an example subject. Circled regions show where the CT showed radiographic change post-RT. We see the result of the over-segmentation in the conventional method and the improvement on this using the proposed method.

FIGURE 9

Summary of percent changes in vessel volume as a function of dose. All groups show increasing reductions in vessel volume with increasing dose however the magnitude of the changes differs in behavior between groups (A) and (B). Group (A) shows minimal change in the unirradiated dose bin while Group (B) shows a large change.

FIGURE 10

Summary of the percent changes in volume of vessel in each of the dose bins for the group (A) swine (all not fed), group (B) swine, and split group (B) swine results masked by being in a fed or not fed region. All data sets show strong linear correlation with dose where the reduction in vascular volume increases with increasing dose. The percent changes in the not fed regions are very similar to the group (A) percent changes while the fed regions show significantly higher magnitudes of change.

FIGURE 11

Change in vessel volume vs change in perfusion. Each point represents a different analysis contour and is the average of the five swine subjects with error bars representing the standard deviations in each metric.