Chapter 3 Automatically Generate a Simplified Chest Atlas from the Chest Computed
3.2 The Experiment Steps
3.2.4 Ordering the Centroid from K-means
After K-mean algorithm is finished, we get all the centroids coordinate. Although the coordinates were gained, we need to order the results so that we can use the results from this ordering to mark the body parts. If you wanted to choose 5 this means that K would be set as five in the K-means algorithm. Than this would gain the centroid coordinates.
This can be used x coordinates to order in 2 dimensional. The smallest x’s coordinates would be the first point. This is ordered from smallest to largest. The last x coordinate would be the biggest, seen in Figure 3.15.
Figure 3.15 Ordering 5 centroid
If there are six body parts, the K value would be six in the K-mean process. It is the same process as mentioned above with arranging the x coordinates from smallest to largest. The third and fourth coordinates would be arranged according to the y coordinates. The biggest coordinates of the third and fourth coordinates would be arranged as the final third coordinate. The smallest y coordinates between the third and
37
fourth coordinates would be the final fourth coordinates. This is seen in Figure 3.16.
Figure 3.16 Ordering 6 centroid
Lastly we arrange the results according to the anatomy’s names. The first point would be arranged to the left upper limb and etc…These body parts are from anatomy books, as seen in Table 3.2. For example, as seen in the image above (Figure 3.15) the yellow number 1’s location, we would confirm it with the table, where 1 is the ‘left upper limb.
38
Table 3.2 The results of body parts according to the ordering points.
My proposed method algorithm is summarized as follows:
Step1: Use the image through recursive erosion, dilation, to eliminate the platform chest image.
Step2: Use Step1’s result, through Sobel edge detection to get binary image.
Step3: Use Step1’s result, through thresholding to obtain an image.
Step4: The results from Step2 and Step3 through AND operation in mathematically morphology, obtain an binary image.
Step5: Step4’s image results, nonzero value pixel’s coordinate as an input, specify a starting a k value, use K-means algorithm to cluster, to obtain clustering results. This result includes the centroid and labeling on the image.
Step6: Use Step 5 of centroid coordinates, and take use of our specified ordering criteria sorting. The result is sorting. Each centroid is referenced from anatomy; use the definitions and labeling the definitions on the images.
39
40
Chapter 4 Experiment and Discussion
In this chapter, we will estimate how useful the methods were. During testing we used NCKU hospital doctors, chest CT file. The useful criteria, we used anatomy [8, 12, 17]
according to the pictures and where the points were marked and compare the images we made by the book’s. The coordinates on the images where we used k-means algorithm are the centroid’s coordinates. Then it was compared to the points of the body parts in the book about anatomy. The comparison is done by using the Euclidean distance [7].
We used a total of ten patient’s CT image to do experiments. Due to the different size of body shape or the difference in position, the patients CT image had different numbers.
Patient 1 has 31 images, patient 2 has 28 images, patient 3 has 28 images, patient 4 has 36 images, patient 5 has 32 images, patient 6 has 35 images, patient 7 has 31 images, patient 8 has 33 images, patient 9 has 27 images and patient 10 has 28 images. All together there were 309 images.
We arranged all the information according to sex and the other group was arranged ignoring the difference in sex. These three groups chest CT images were get five to six body parts. According to the above Euclidean distance, we calculated our location of the body parts in different images to the book. Due to the difference images, some body parts were not clear or were not taken. So if we estimated the location of these unclear body parts, the results would not be evident.
All experiments were performed on a Intel® Core™ 2 Duo CPU E8300 @ 2.83GHz personal computer with 2 GB main memory and 250 GB hard disk running on Windows XP Service Pack 3.
In section 4.1, we will compare and show the images for using proposed method.
41
Pat
1 2 n
4.1 The Results Using the Proposed Method to Ten ients
4.1.1 The Results Using the Proposed Method
Before going into the subject we will introduce Euclidean distance. In the area of mathematics, the definition of the Euclidean distance or otherwise known as the Euclidean metric is the ordinary distance between two points. These two points are often measured with a ruler. This can be proven by repeating by using the Pythagorean theorem.
Using this formula as a distance, Euclidean space will turn into a metric space. The associated norm is called the Euclidean norm.
Previous literature says that this metric as Pythagorean metric. However, it is important to note that this technique has been rediscovered numerous times in history and it has a logical extension of the Pythagorean Theorem.
The Euclidean distance between points U =( ,u u ,...,u ) and V =( ,v v1 2,... )vn , in Euclidean n-space, is defined as
For example, for two 2D points, U= and V= , the distance is computed as:
1 2
( ,u u ) ( ,v v1 2)
2 2
1 1 2 2
(u −v) +(u − )v (4-2)
In this research paper we use Euclidean distance were all on 2D data, similar to the above chapter that introduced K-means algorithm. In this algorithm, we calculated the sample points and the centroid’s distance. This is done by using the Euclidean distance.
Than this is one main criterion in accessing if the results from this experiment.
4.1.2 The Results Using the Proposed Method in Choosing Six
Body Parts.
In this section, we use the original CT to automatically get six body parts to be labeled with organ definition. These six labels are body of vertebra, spinal canal, right and left upper limb, left and right subscapularis muscle or head of rib. In this section, due to the enormous data image so we picked images from patient 1. The images chosen were the best case and the worst case to display and describe.
42
43
(a)
(b)
(b)
Figure 4.1 Best case of six body parts chosen (a) Original (b) Labeled image
In Figure 4.1, it is the most typical CT image that we want to process. The image of bones and organs are clearly visible. The results of this typical image would be fair.
Lastly through calculating our ideal coordinates, the differences were not bad. The
Euclidean distance was only 4.0181.
(a)
44
(b)
(b)
Figure 4.2 Worst case in choosing six body parts. (a)Original image (b)Image after labeling
As seen in Figure 4.2, the first is the worst case of six body parts chosen. This picture
clearly shows that the points chosen are invalid. This is not what we had expected. The Euclidean distance mean is 42.0539. The cause of this is simply that the body of vertebra was not taken clearly. It’s pixel’s gray level and the surrounding muscle and tissue’s gray level is about the same. As seen in Figure 4.2(a) where the red arrow head points.
4.1.3 The Results Using Proposed Method in Choosing Five
Parts.
Body
In this section, we will use the original CT to automatically five body parts. After that the definitions of human organs are labeled. The five labels are spinal canal, right and left upper limbs, right and left subscapularis muscle or head of rib. This can be seen in Figure 4.3.
(a)
45
(b)
Figure 4.3 Best case in choosing five body parts (a) Original image (b) Image after labeling
As seen in Figure 4.3 it shows the best case of choosing five body parts. The red arrow head in Figure 4.3(a) shows the body of vertebra. The body of vertebra in this image was not taken clearly. It will be eliminated by thresholding because it wasn’t taken clearly. This would not affect the result of choosing five body parts. Thus the five body parts chosen are much more accurate. This last image’s Euclidean distance mean is 3.4748.
46
47
(a)
(b)
Figure 4.4 Worst case in choosing five body parts (a) Original image (b) Image after labeling
As seen in Figure 4.4 it is the worst case in choosing five body parts. It is the opposite of the image above, where the body of vertebra is taken clearly. In Figure 4.4(b) the yellow arrow head shows that. The spinal canal’s original location is marked by the
48
red arrow head. Due to the K-means clustering was influenced by pixel around the body of vertebra. The chosen spinal canal will be deviated upwards. This would also deviate other chosen body parts. This image’s Euclidean distance mean is 37.6147.
4.2 Special Case: The Invasion of Metallic Items
According to the patient’s chest CT, during the experiments there will be numerous special images that would appear. These would influence the results of the experiments .Special situations are introduced below.
During experiments, some numbers were not acceptable, so images were closely monitored to see what the problem was. We discovered some CT images had human tissues and some non human tissues. These non human tissues were like metallic items or pipes that were inserted for medical use. This situation is seen in patient 8 and patient 9.
(a)
(b)
49
(c)
Figure 4.5 An unknown item invasion in Patient 8 (a) Original (b) Results from choosing 5 body part. (c) Results from choosing 6 body parts.
As seen in Figure 4.5(a) the red arrow head shows an invasion of an unknown item.
Due to unable to eliminate the unknown item, this influenced the results greatly.
Choosing five body part’s Euclidean distance mean were 45.0139 and the body parts where 6 were chosen had a Euclidean distance mean of 23.5105. There were some images that had unknown items invasion, had a mean over 100, this could influence greatly on the results.
50
Table 4.1 Obtaining five body parts, using pixel as an unit to calculate Euclidean
Patient1 3.4748 37.6147 13.1242 9.2494 Patient2 4.3778 40.4950 23.6917 13.4649 Patient3 2.7944 42.1344 26.2717 15.1543 Patient4 6.2081 45.7356 27.0521 16.5821 Patient5 3.3231 39.2526 15.7397 11.7670 Patient6 4.6041 47.3880 33.9643 12.2842 Patient7 6.3187 54.8082 40.6728 9.9252 Patient8 4.5551 125.0787 41.1615 38.9842 Patient9 22.7687 43.0277 33.7533 5.3916 Patient10 27.2089 45.3313 38.5092 3.5560
Male 4.5551 125.0787 35.6390 19.5601
Female 2.7944 42.1344 19.4075 13.4802
All patients
2.7944 125.0787 29.3781 19.12263
51
Table 4.2 Obtaining six body parts, using pixel as an unit to calculate Euclidean distance.
Patient1 4.0181 42.0539 21.9965 11.3501
Patient2 3.1319 12.2472 6.6272 2.4241
Patient3 3.7466 16.0136 7.7499 3.3473
Patient4 4.6648 21.8054 7.8267 2.9325
Patient5 4.5572 39.3277 15.4894 10.8374
Patient6 4.8097 14.6193 7.8810 2.6144
Patient7 3.0581 16.9723 9.9050 5.1280
Patient8 4.7910 119.4405 35.2707 39.4599 Patient9 3.9855 46.4590 13.9921 13.4462 Patient10 3.8667 10.9022 6.5793 1.5851
Male 3.0581 119.4405 13.6347 19.9956
Female 3.1319 42.0539 13.2783 10.3227
All patients
3.0581 119.4405 13.6354 17.1642
52
Table 4.3 Obtaining five body parts and eliminating the special case.
Male 4.6041 21.8054 35.6390 19.5601
Female 2.7944 42.1344 19.4075 13.4802
All patients
2.7944 42.1344 27.3544 15.2300
Table 4.4Obtaining six body parts and eliminating the special case.
Best
Male 3.0581 16.9723 8.0682 3.4776
Female 3.1319 42.0539 13.2783 10.3227
All patients
3.0581 42.0539 10.5582 7.9870
53
4.3 Conclusion
Figure 4.6 Patient 1 choosing six body part’s Euclidean distance histogram.
Figure 4.7 Patient 1 choosing five body part’s Euclidean distance histogram.
Lastly, we will use the image’s data to organize into tables as seen in Table 4.1 to Table 4.4. Table 4.1 represents the results of five body parts chosen and its average. Table 4.2 represents results of six body parts chosen and its average. Figure 4.6 and Figure 4.7 represent patient 1 choosing five and six body part’s Euclidean distance histogram.
54
55
Horizontal axis represents the number of images; the vertical axis represents the Euclidean distance. There are total 10 patient’s individual’s best and worst case and each individual’s average. The best and worst case is in each image’s five or six marks and the ideal five or six’s marks Euclidean distance’s mean and this mean is the biggest or the smallest. The data is also separate into female and males.
Data analysis shows that most chest CT images, six body parts are much accurate.
However, there are some patient’s images that some muscles or organs were not taken clearly. Most of these examples are body of vertebra was not taken clearly, this shows it would be best if five body parts are chosen. As seen in Figure 4.6 and 4.7, arrowheads represents 31 image, due to the body of vertebra was not taken clearly, so choosing five body parts is much more accurate than in choosing six.
Out of the 10 patients, Patient 8 and Patient 9 show outside invasions. After analyzing the results, outside invasions could probably is medical intravenous drip or medical instruments used in aiding the patient. This would affect the results. Eliminate patient 8 and nine to calculate the results as seen in Table 4.3 and Table 4.4. If eliminating patent 8 and 9 the average would much more acceptable. Thus, if five body parts are chosen, its Euclidean distance’s mean is 27.3544. If six body parts are chosen, its Euclidean distance’s mean is 10.5582. These results are much more acceptable.
It is important to note that the patient’s sex and our method also influence our data as seen in Table 4.3 and Table 4.4.
56
Chapter 5 Conclusion and Future Work
5.1 Conclusions
With unsteady improvement of technology, lifestyles of humans are getting more and more improvements. With industrialization shaping our society life, more and more diseases are becoming often visible. With improvements of food, the nutrients are lost during the improvements and the mass of pollution from industries, the chances of getting cancer is growing at a high rate. Although technology is improving, judgment in medical fields are still using X rays or CT to help doctor’s judgments.
In this paper, a simple method is introduced to automatically analysis chest CT and locating each organ to an improved atlas.
5.2 Future Works
The result accomplished in this thesis is only a preliminary of the image registration.
Much of work is needed to be done in the future. Some of the future works are:
(1) Further atlas
In this paper, due to the lack of ability and time, CT is only concentrated on six locations on the chest or organs. Using this method, more research can be conducted on more different areas.
(2) Choosing Chest photos and getting the automatic K value.
In this paper, it’s an automatically producing CT atlas, however, some parts is done
57
by human power. This is seen in choosing chest’s CT, using anatomy books as references to choose chest CT tools. Using K-means mathematically, automatically getting the K value [3] done with the photographs. These are not automatically done but done by humans.
(3) Three dimensionally images, Whole body atlas
In this thesis, the images we use are 2D. However, human organs are 3D. This would be lost as the images are only 2D. The properties will be lost during the transformation.
Wishing that juniors could transform the 2D into 3D [18, 21, 27, 29] for further research, and produce the whole body atlas. These atlases could be given to medical staffs for future judgments.
58
[12] 陳金山、徐淑媛編譯，L. H. Blackbourne, J. Antevil and C. Moore 著，2001，回 憶解剖學，195~199，台北:藝軒圖書出版社。
[13] 張豫雯，民 92，應用資料探勘挖掘潛在電子商務客戶群，大同大學資訊工程學 系碩士論文。
59
[20] A. T. Joseph, “Understanding the Standerdized Uptake Value, Its Methods, and Implications for Usage,” The Journal of Nuclear Medicine, Vol. 45, No. 9, pp.
1431-1434, 2004.
[21] G. Li, T. Liu, G. Young, L. Guo and S. T. C. Wong. “Deformation invariant attribute vector for 3D image registration: method and validation,” Biomedical Imaging:
Nano to Macro, 2006. in 3rd IEEE International Symposium, pp. 442-445, 2006.
[22] J. Y. Huang, P. F. Kao and Y. S. Chen, “A Set of Image Processing Algorithms for Computer-Aided Diagnosis in Nuclear Medicine Whole Body Bone Scan Image,”
IEEE Trans. on Nuclear Science, Vol. 54, No. 3, 2007.
[23] J. Serra, “Image Analysis and Mathematical Morphology,” Academic Press, London, 1982.
[24] M. Gudmundsson, E. A. E1-Kware and M. R. Kabuka, ＂Edge Detection in Medical Images Using a Genetic Algorithm Future,” IEEE Trans. on Medical Image, Vol.17, No.3, 1995.
60
[25] N. Otsu, “A Threshold Selection Method from Gray-Level Histograms,” IEEE Transactions on Systems, Man, and Cybernetics, Vol. 9, No. 1, pp. 62-66, 1979.
[26] O. Demirkaya, “ Lesion Segmentation in Whole-Body Images of PET,” 2003 IEEE Nuclear Science Symposium Conference Record, Vol. 4, pp. 2873-2876, 2003.
[27] P. J. Besl and N. D. McKay, “A Method for Registration of 3-D Shapes,” IEEE Trans.
Patt. Anal. Machine Intell., Vol. 14, No. 2, pp. 239-256, 1992.
[28] S. Chen and R. M. Haralick, “Recursive Erosion, Dilation, Opening, and Closing Transforms,” IEEE Trans. Image Processing, Vol. 4, No. 3, 1995.
[29] S. Gefen, L. Bertrand, N. Kiryati, and J. Nissanov, “Localization of Sections Within the Brain Via 2D to 3D Image Registration,” Acoustics, Speech, and Signal Processing, 2005. Proceedings. IEEE International Conference, Vol. 2, pp.733-736, 2005.
[30] T. K. Yin and N. T. Chiu, “A computer-aided diagnosis for locating abnormalities in bone scintigraphy by a fuzzy system with a threestep inimization approach,”
IEEE Trans. Med. Imag. , vol. 23, pp. 639–654, 2004.
[31] R. C. Gonzalez and R. E. Woods, Digital Image Processing Second Edition. Prentice Hall, 2002.
[32] R. C. Gonzalez and R.E. Woods, Digital Image Processing Third Edition. Prentice Hall, 2008.
[33] R. M. Haralick, S. R. Stemberg, and X. Zhuang , “Image analysis using mathematical morphology,” IEEE Trans. Pattern Anal. Machine Intell. , Vol. 9, No.
4, pp. 532-550, 1987.