Discussion and Conclusions

In this section, even though experimental results demonstrate that the QHCF leads to better segmentation results over other approaches, including UGHCF and adaptive ICM, same as what occurred in other random field based solutions, the determination of parameters including Tc, P1,P2, and Tnin is also a difficult part of implementing the QHCF into real applications. This is mainly because the evaluation of segmentation result is usually application oriented, which highly depends on the verifications of object identification/recognition at the higher level in the system. Therefore, it is hard to find an ideal measurement in the lower level to feedback the segmentation performance. Even though there is no theoretical approach on the automatic optimal searching, some heuristic solutions can be adopted in our implementation.

(i) Application-oriented empirical parameter selection: For different applications, the requirements of segmentation results may be different

Figure 8.5: Comparison of initialization sensitivity between UGHCF and QHCF. The original is in Fig. 7(a). (a)-(c) Segmentation results under uniform grid initialization with grid size 10, 20, and 30 pixels, respectively. (d)-(f) Results Reliability Criterion threshold Trc = 5, 10, and 20, respectively. The result within the white circle demonstrates the stability difference of the two algorithms with different initialization.

Figure 8.5: Comparison of initialization sensitivity between UGHCF and QHCF. The original is in Fig. 7(a). (a)-(c) Segmentation results under uniform grid initialization with grid size 10, 20, and 30 pixels, respectively. (d)-(f) Results Reliability Criterion threshold Trc = 5, 10, and 20, respectively. The result within the white circle demonstrates the stability difference of the two algorithms with different initialization.

even with the same input image. Therefore, for a specific type of images, some empirical selections of parameters can be adopted. The parameters for two categories of images have been analyzed: one is about lumen segmentation with T1W MR images; the other is about the frame segmentation in videoconference clip. Table 8.1 shows the typical values of parameters in two applications for the QHCF algorithm.

Figure 8.6 is an example of the segmentation with different parameter combinations on T1W MR images. It shows that the parameter combination Trc = 10, Tnm = 10, P1 = 600, and ยก32 = 100 has better performance

Table 8.1: Parameters for QHCF algorithm

Empirical T1W MR Video conference frame parameters Image (QCIF)

ft ft

10 600 1000 10

10 400 600 10

10 600 1000 10

10 400 600 10

Figure 8.6: Segmentation of carotid artery lumen by QHCF, image size: 90 by 90, Trc = 10, Tmin = 10. (a) Original image. (b) Canny edge map. (c) 01 = 600, 02 = 600. (d) 01 = 600, 02 = 800. (e) 01 = 600, 02 = 1000. (f) 01 = 600, 02 = 1200. (g) ft = 400, 02 = 1000. (h) 01 = 800,02 = 1000. (g) 01 = 600, 02 = 1000, Tmin = 30.

Figure 8.6: Segmentation of carotid artery lumen by QHCF, image size: 90 by 90, Trc = 10, Tmin = 10. (a) Original image. (b) Canny edge map. (c) 01 = 600, 02 = 600. (d) 01 = 600, 02 = 800. (e) 01 = 600, 02 = 1000. (f) 01 = 600, 02 = 1200. (g) ft = 400, 02 = 1000. (h) 01 = 800,02 = 1000. (g) 01 = 600, 02 = 1000, Tmin = 30.

than others for lumen segmentation because all the typical regions, including lumens and blood vessel wall, are partitioned correctly. To further fine-tune the results, we increase the minimum region threshold as Tmin = 30 and an even "clear" result can be obtained as shown in Fig. 8.6(g).

(ii) Supervised segmentation with interactive parameter selection: Another solution for the parameter estimation in implementing QHCF algorithm into real application is to provide an interactive parameter selection interface for users. In our implementation, two modes were designed: auto and manual. In auto mode, a few empirical parameter combinations for typical applications are stored. The user can start with these predefined values and find one that can generate satisfactory results. Any further fine-tune of the segmentation performance can be achieved by switching to manual mode in which the parameters can be further adjusted.

0 0

Post a comment