Modern Computed Tomography technology enables entire scans of the lung with submillimeter voxel precision. However, manual segmentation of the complex vessel tree structure is not only an extenuatingly long task for the human; it can also be considered an almost impossible mission for several reasons: first, the boundaries of a vessel (especially the thin ones) are quite difficult to determine in a consistent way due to image noise; second, manual measurements of the vessels in order to assess diameters and branching angles via 2D images are unreliable, mainly because of problems of foreshortening in projected view. Therefore, highly automated segmentation of the pulmonary vascular tree based on 3D image analysis plays a fundamental role in detecting and characterizing the vessel structure.
Automated lung vessels segmentation
As a consequence, the automated segmentation of the vessel tree structure of the lung was understandably considered by our client a crucial process in order to obtain an optimal interpretation of pulmonary CT scans.
The lung vessel segmentation solution which we provided to our client focused on a complex set of techniques, the core of which is centered on the concept of strain energy vesselness and the use of vessel enhancement filters.
These filters derive structural information from a Hessian matrix computed using Gaussian derivatives. The role of the matrix is to help identify local forms in a specific region of interest. In this case, the targets of the lung vessel enhancement are cross-sections of tubular objects, the round forms of which correspond to the location of blood vessels.
Three invariants are then measured on the values of the Hessian matrix: intensity contrast, structure strength and shape. The purpose of these values is to exclude undesired forms (like lung boundaries) from the image structure and to identify continuous forms bearing consistent intensity to locate the vessel which compose the extracted vascular tree. In addition, the method has also helped the client to identify sparse or void regions of the vascular tree, which correspond to the horizontal and the oblique fissures of the lung.
We intentionally decided to test our method against a set of data which presents statistical challenges: respiratory diseases like emphysema, pulmonary nodules or embolisms affect the lungs in a way that introduces additional difficulties to the lung vessel segmentation. The method has proved practical and reliable: the results show that this technique of automatic pulmonary vessel segmentation completes very effectively the vessel tree structure provided by the CT scans of the lung, in such a way that the resulting image is more precise and matchlessly faster than any tentative manual segmentation could be.