Automated Main Arterial Region Separation in Brain MR Angiography for Improved Clarity and Efficiency in the Clinical Visualization Process
 
Authors:
Peter J. Schommer, MSEE, IBM; Nelson Ramirez; Xiaojiang Yang, PhD; Daniel J. Blezek, PhD; Janice R. Glowacki; William J. Ryan
 
Background:
With ever increasing demands on radiology departments, efficiency improvements in the clinical workflow via computer automation of manual processes will serve to improve patient care and reduce health care costs. Intracranial time-of-flight (TOF) MRA examinations can be viewed using a Maximum Intensity Projection (MIP) rendering. In MIP renderings containing many vessels, some vessels obscure the view of others, obscuring clinically relevant pathology. By independently rendering three vascular regions, clarity is increased by removing obstructing vascular structures. Mayo Clinic technicians currently perform this process manually, separating the TOF into the basilar, right carotid and left carotid arterial regions. The manual separation process takes ten minutes or more per patient. The goal of this work is to develop an algorithm automating the process, thereby reducing technicians’ workload, and to improve workflow efficiency by freeing the technicians to perform other tasks during the time normally dedicated to this process.
 
Evaluation:
With proper Institutional Review Board approval, we de-identified 324 retrospective TOF series for use in developing the algorithms. These exams have the technician-generated MIP renderings of the 3 vascular regions.

The algorithm automatically identifies the three arterial regions, based on a combination of information derived from an atlas and patient-specific vasculature structure. The atlas is registered to the patient data and used to identify the general location of the vascular regions[1]. Patient-specific vasculature is extracted through vessel segmentation, and techniques from graph theory classify each segment into one of the three vascular regions[2]. The results of both approaches are combined to produce a final resulting volume for MIP rendering.

Atlas Construction:
To construct our atlas, the technician generated MIP renderings from 84 TOF datasets were backprojected into the TOF. Each voxel had three labels corresponding to each of the three vascular regions. The TOF images from each patient were registered into a common coordinate system using an affine transformation. Four atlases were constructed (Figure 1): a) an anatomic atlas, which is an average of all the TOF images; and b) probabilistic atlas of the right carotid, basilar, and left carotid regions, consisting of the probability that each voxel was contained in that region.
 
Figure 1
Figure 1. Atlases: (a) Bottom MIP of the anatomical atlas ( average intensity atlas ) from 84 TOF datasets;
(b) Side view of probabilistic atlases from 84 TOF datasets using gradient shading.
 

Using the Atlas:
The three vascular regions on a new TOF study are identified by the following steps: 1) the anatomic atlas is registered to the new TOF study using an affine registration process; and 2) the three probabilistic atlases are transformed by the transform parameters from the anatomic atlas registration. Once the atlases and the patient data are in the same coordinate system, they are compared on a per voxel level. The probabilistic atlases only provide information as to the general location where the basilar, left carotid, and right carotid regions should be located. For peripheral vasculature this information is usually correct. However, around the circle of Willis, the atlas information is not accurate, due to the large variability in patient anatomy. To provide accurate separation in these regions, a patient-specific approach is used and described below.

Vessel Segmentation and Graph Construction:
Vessel segmentation is obtained using region growing[3]. A graph is constructed from the centerline of the segmented vasculature. The nodes of the graph correspond to vessel junctions, and the edges are the connecting vascular segments. Using the graph representation, spatial interconnections between vessel segments are analyzed. The graph is overlaid on each of the three probabilistic atlases. A set of features is generated for each graph edge including the average radius, length, centroid, and average overlap with the basilar, right carotid, and left carotid atlas.

 

Figure 2
Figure 2. Automated Process: (a) A new TOF volume; (b) Bottom MIP view of the anatomical atlas ( average intensity atlas ) from 84 TOF datasets; (c) Vessel Segmentation; (d) Graphs generated from segmentation;
(e) Side view of probabilistic atlases from 84 TOF datasets using gradient shading.
 
After graph generation and graph edge feature production, we classify each vessel edge into 7 main categories. Each piece of vessel can be entirely in the basilar, right carotid, left carotid, or in any combination of these three regions. Each edge is classified based on heuristics and graph pattern searching approaches. Figure 2 demonstrates the process flows for generating the graph analysis and atlas classifications.
 
Figure 3

Figure 3. Combining atlas and vessel classification results: (a) Centerline graphs; (b) Side view of probabilistic atlases; (c) A new TOF; (d) Side view of probabilistic atlases with additions/deletions according to classified vasculature.
 

Combining Results:
After graph edges have been classified, the probabilistic atlas regions are merged with the information from the classified graph edges to add vessels that were missed by the atlas or incorrectly removed from each vascular region. For example, if the basilar region of the atlas does not cover the true basilar vessels completely, the segmented vessels that are classified as basilar, together with a 3D neighborhood around the vessel edge, will be added to the basilar region that would be removed from the source TOF as shown in figure 3. The approaches complement each other. The outer regions cannot be separated using patient specific vasculature via vessel segmentation, and the central regions cannot be separated by an atlas alone. The combination provides the best of both population and patient specific approaches. Figure 4 shows a comparison of the automatic versus manual brain vascular region separation results.

 

Figure 4
Figure 4: Comparison of automatic versus manual separation results: (a) right carotid automatic; (b) right carotid manual; (c) left carotid automatic; (d) left carotid manual; (e) basilar automatic; (f) basilar manual.
 
Discussion:
The algorithm averages 3 to 4 minutes to run per volume on a modern multi-core x86 workstation. The output is 3 new volumes corresponding to patient data in the right carotid, left carotid, and basilar regions with all other voxels set to zero. The segmented volumes are MIP rendered to produce the final output. Validation was a visual inspection under the guidance of a subspecialty radiologist. Based on this initial evaluation of 324 cases, 74% are visually close enough to the technician drawn results to be placed in an acceptable category; 18% have minor problems but are likely acceptable; and 8% have more severe problems where clinical acceptability would be unlikely. We have developed a display application to allow blinded evaluation by 3 radiologists, and will have these additional results within the month. The main potential improvement may be achieved by better separation of vasculatures that deviate significantly from the atlases. Composed of 84 datasets, our atlas encapsulates a fair amount of patient variability. Further improvements are likely possible using additional graph analysis techniques such as semi-supervised graph labeling algorithms[4]. Using both an anatomical atlas and patient-specific information from vascular segmentation allows the algorithm to handle both peripheral regions and the fine-level separation required in the central vasculature.
 
Conclusion:
The algorithm holds promise toward automating the technician’s task of generating MIP renderings of each of the three vascular trees. The techniques developed in this system are applicable to MRA exams of other anatomy such as the neck.
 
References:

[1] Ibáñez L, Schroeder W, Ng L, Cates J. The ITK Software Guide. Second Edition. Updated for ITK version 2.4. Insight Software Consortium. Kitware, Inc. http://www.itk.org. November 2005.

[2] Uchiyama Y, Yamauchi M, Ando H, et al. Automated Classification of Cerebral Arteries in MRA Images and Its Application to Maximum Intensity Projection. Engineering in Medicine and Biology Society. 28th Annual International Conference of the IEEE Aug. 30-Sept. 3. 2006;4865 - 4868.

[3] Yang X, Blezek DJ, Cheng LT, Ryan WJ, Erickson BJ. Computer-Aided Detection (CAD) of Intracranial Aneurysms in MR Angiography. Presented at the SIIM 2009 Annual Meeting. Charlotte, NC. June 2009.

[4] Zhou D, Bousquet O, Lal TN, Weston J, Schölkopf B, Olkopf BS. Learning with Local and Global Consistency. Advances in Neural Information Processing Systems 16. MIT Press. 2004;321-328.