Xiaojiang Yang, PhD, Mayo Clinic; Daniel J. Blezek, PhD; Lionel T. Cheng, MBBS; William J. Ryan; Bradley J. Erickson, MD, PhD, FSIIM

We hypothesize that an algorithm which highlights features suggestive of aneurysms on intracranial time-of-flight (TOF) MRA examinations can be developed. This, in turn, will greatly assist radiologists in detecting intracranial aneurysms.

Detecting intracranial aneurysms is very important in preventing severe morbidity or death. To detect intracranial aneurysms, non-invasive 3D TOF MRA is often used as the substitution for generally considered gold standard, digital subtraction angiography (DSA), which is an invasive method. A typical 3D TOF MRA dataset usually contains more than 100 images, making it very time- and labor- consuming for radiologists to detect aneurysms.

In recent years, efforts in computer-aided detection (CAD) of aneurysms have been made by many researchers to help radiologists find aneurysms from MRA data sets.[1,2,3] Intracranial aneurysms can be categorized, according to shapes, into three types based on morphology: saccular, bifurcation, fusiform; and, according to sizes: big or small (5mm diameter is a common breakpoint). Until now, none the proposed methods could detect all these types of aneurysms. In addition, some methods are not fully automatic, which means human intervention is required.

For example, Arimura described a scheme that is prone to missing both large aneurysms and saccular type small aneurysms.[1] Hasanori reported a simple method they developed, which is not fully automatic, and can not detect small aneurysms.[2] Kobashi proposed a method but it cannot find fusiform aneurysms.[3]

In our research, the goal was use of a fully automatic scheme to detect all types of unruptured and untreated aneurysms from 3D TOF MRA data sets. Experiments showed that the scheme developed reached our goal with an acceptable average false positive (FP) number.

After IRB approval, we identified 58 3D TOF MRA studies in patients who had undergone intracranial DSA to confirm the presence of 1 or more aneurysms. These studies were then annotated by a trained radiologist to identify the aneurysm(s). The radiologist had access to reports and DSA images to increase confidence of findings.

The algorithm we applied consists of the following steps.
1. A Gaussian filter is applied to the original 3D MRA data set to denoise the images.

2. The resulting 3D MRA data set is rescaled to be isotropic using the linear interpolation method.

3. An automatic segmentation algorithm is used. The result is one or multiple separate 3D regions. The detection procedure will be applied to each of those 3D regions.

4. The center lines of the 3D regions are calculated using a 3D thinning algorithm.[4] Along each center line (we call trunk), two center point features, radius and radius remainder, are calculated for every point. A radius remainder represents the roundness of a vessel around a center point. Trunk-level features, like length, minimum radius, maximum radius, average radius, are also calculated for every trunk.

5. Points of interest (POIs) are generated. We collect four types of POIs (Figure 1): (a) short branch, (b) bifurcation, (c) local maximum, and (d) cyclic short trunk. These types of POIs can cover all shapes and sizes of aneurysms. Among these types, the first three come from the surface of vessels, but the last one comes from the center line. POIs of type (c) are calculated via a single point seeded distance transformation [5], and all other POIs are obtained by analyzing trunks. A group of features for all (a)-(c) POIs are then calculated, including distance to the center line, features of the nearest point in center line, by features of the nearest center line, and by plainness and cylinder surfaceness of a vessel in the POI point. The plainness and cylinder surfaceness characterize the normalness of a straight vessel in a specific point.

Figure 1: Four types of POIs

6. Several sieving strategies are applied to the POIs to remove FPs.

7. A clustering technology is used to further remove FPs.

The final output of aneurysm suspects are then represented by a translated point set, which contains the middle points of the suspect POIs and their nearest points in the center line.

We executed the algorithm on each of the 58 examinations, and computed the number of TPs as those where POIs were within 10mm of the annotated aneurysm location. FPs were all the POIs more than 10mm from the annotation. Since the purpose of the annotation was to highlight suspected regions for purpose of visualization of that region by a radiologist, we did not require exact overlap.

58 randomly chosen 3D TOF MRA data sets from Mayo Clinic were used to test our scheme. These studies contained 21 aneurysms, including all the three shapes, with the minimum size 2mm and maximum size 33mm.

Initially in our test, 18 aneurysms were successfully detected fully automatically, with an average FPs of 4.84 per examination. We found that the missing 3 aneurysms were due to the failure of segmentation that missed the aneurysms. If we manually adjusted the threshold used to segment the vessels in those three cases, all 21 aneurysms could be detected, with an average FPs 4.8.

Intracranial aneurysms represent a cause of morbidity and mortality that is difficult to detect. There are many factors that can help identify at-risk populations for aneurysm, increasing the number of MRA screening studies. However, even for experienced radiologists, the sensitivity to 3~7mm aneurysms is only about 40~50%. The addition of a CAD scheme, which could highlight suspicious regions, while have a low FP rate, can improve the quality of health care provided at little or no increased radiologist effort. The result of the algorithm we developed reflects a high accuracy when vessels have been properly segmented. We believe not only that 5 FPs per case is in the acceptable range, but also that the number of FPs can be further reduced by combining some machine learning algorithms into the sieving process.

The CAD scheme developed in this study is fully automatic. It can detect all shapes of aneurysms of different sizes with high accuracy. We believe it will be a useful tool assisting radiologists in the detection of aneurysms.

[1] Arimura H. et al. Automated Computerized Scheme for Detection of Unruptured Intracranial Aneurysms in Three Dimensional Magnetic Resonance Angiography. Academic Radiology. October 2004;Vol.11:10.
[2] Hisanori H, et al. Development of Cerebral Aneurysm Computer-Aided Detection Systems with 3D MRA Data. Yokogawa Technical Report English Edition. 2005;39. Available at: http://www.yokogawa.com/rd/pdf/TR/rd-tr-r00039-008.pdf.
[3] Kobashi S, et al. Computer-Aided Diagnosis of Intracranial Aneurysms in MRA Images with Case-Based reasoning. IEICE Trans. Inf. & Syst. January 2006;Vol.E890D:1.
[4] Lee Ta-Chih, et al. Building Skeleton Models via 3-D Medial Surface/Axis Thinning Algorithms, CVGIP: Graphical Models and Image Processing. November 1994;Vol.56:6:462-478.
[5] Zhou Y, et al. Efficient Skeletonization of Volumetric Objects, IEEE Transactions on Visualization and Computer Graphics. 1999;Vol.5:3:196-209.