D

Figure 8.10: An example of lumen segmentation with MRF-based active contour framework. (a) The original MR image. (b) Segmentation results with adaptive ICM. (c) Segmentation results with the QHCF algorithm. (d) Rough lumen contour based on the QHCF segmentation result. (e) The six selected control points for MPA model initialization. (f) The fine-tuned lumen contour achieved by applying the MPA algorithm. In comparing (d) and (f), it becomes clear that contour tracking under the proposed framework results in superior smoothness control than with the MRF-based solution.

sequence. It is normally impossible to know where this kind of topological transformation happens.

8.3.4.2.1 Lumen Region Identification. To address the problem of topographical changes, we integrated prior knowledge into the control points searching procedure. The bulk of our work is represented by the second block in the processing diagram shown in Fig. 8.7.

First, we model the MR image with MRF model and segment each of them into many regions by applying the QHCF algorithm. Since the number of lumen region may vary due to the bifurcation of carotid artery along the image sequence, a lumen identification process is indispensable before further plaque analysis. For each image, the lumen identification is achieved by letting all the segmented regions through a knowledge-based decision tree and picking up lumen region(s) of interest. The decision criteria are obtained by analyzing the statistical distribution of lumen region features based on prior knowledge in the test dataset. In the atherosclerotic blood vessel study, the following features are regarded critical for lumen identification:

(1) Region area CArea

(2) Region average intensity CIntensity = N^Z^ (8.37)

(3) Region circularity Corcular = LCArea, (8.38)

L Contour where LContour is the length of region contour;

(4) Region location C^cation

The basic structure of the decision tree is shown in Fig. 8.11. For criteria CArea, CIntensity, and CCircularity, statistical analysis of training MR image data required the use of two standard deviations as the satisfactory scale to make sure most of the variation range can be covered. For CLocation, it reflects the maximum radius of lumen center may locate in current slice away from the center of lumen in the previous slice. To reduce the computation in the identification process, the most distinctive feature of the target region is always analyzed first so as to decrease the number of candidates in the following criteria checking. In our study, the sequence is arranged as CArea, CIntensity, CLocation, and

^Circularity.

From above identification procedure, it can be seen that the accuracy of the low-level region segmentation plays a very important role in the topography detection. This can be achieved by using the QHCF algorithm. For the lumen

REJECTED ACCEPTED

Figure 8.11: Diagram of the decision tree structure for lumen identification in MR image sequences.

identification step, however, if the choice of training dataset is sufficiently representative, those criteria will be fairly precise and hence make the decision result more stable. To further enhance the lumen tracking ability along image sequence, the location correlation of lumen regions between adjacent MR slices can be constrained by CLocation, which can further limit the searching range and therefore reduce computation.

8.3.4.2.2 Experiments and Discussion. In this study, 20 MR image sequences were tested with the proposed method. These images were scanned by a 1.5T SIGNA scanner (GE Medical Systems, 5.7 Echo Speed, custom made phased-array coils) in two imaging contrast weightings: T1-weighted (T1W) and 3D time-of-flight (3D TOF). Each mode produces 10 sequences with the bifurcation inside and 12 slices in each sequence.

It is well known that some feature characteristics appear altered by different modalities. The T1W sequence produces a lower intensity of the lumen signal. In 3D-TOF images, the intensity of the lumen is both higher and more uniform because of flow enhancement during imaging. For each contrast weighting, five sequences were selected for identification criteria estimation and the rest were used as testing data. Table 8.2 shows estimated criteria. The error rate is defined as the ratio between number of regions with error detection (misdetection or false alarm) and total number of lumen regions.

Segmentation Issues in Carotid Artery Atherosclerotic Plaque Table 8.2: Estimated criteria for decision tree

3DTOF

Mean

Mean

Error rate

0.23

Error rate

® For the common carotid artery.

For lumen contour tracking, prior use of the MPA model has provided more satisfactory results than MRF segmentation alone. Figure 8.12 is an example of the lumen segmentation procedure. A typical carotid artery MR image is shown in Fig. 8.12(a). Because of noise and artifacts during the imaging process, the intensity of the lumen area is not uniform and contains isolated bright spots. The QHCF algorithm was applied to this image with the result shown in Fig. 8.12(c) (note that lumen segmentation was not affected by bright spots in the lumen or background noise). The contour of interest region is shown in Fig. 8.12(d). It is obvious that some sharp corners on the left part of the contour and the bottom part are also not very smooth. Figure 8.12(e) shows the control points and the final MPA fine-tuned contour is demonstrated in Fig. 8.12(f). Figure 8.13 is an example of the blood vessel tracking in MR image sequence, with lumen bifurcation included.

Even though experimental results demonstrate good performance of the proposed framework, by analyzing those cases with error lumen identification it is found that the decision tree needs to be further enhanced so as to overcome the disturbance caused by random imaging artifacts in lumen region. Moreover, additional criteria and optimal decision strategies should also be considered in future research.

0 0

Post a comment