• No results found

Medical Image Segmentation of PACS System Image Post-processing *

N/A
N/A
Protected

Academic year: 2021

Share "Medical Image Segmentation of PACS System Image Post-processing *"

Copied!
6
0
0

Loading.... (view fulltext now)

Full text

(1)

Y. Zhang (Ed.): Future Communication, Computing, Control and Management, LNEE 141, pp. 495–500. springerlink.com © Springer-Verlag Berlin Heidelberg 2012

Post-processing

Lv Jie, Xiong Chun-rong, and Xie Miao

Department of Professional Technical Institute, Yulin Normal University, Yulin Guangxi 537000, China

lvjie2227@126.com, xiongchunr@163.com, mm.xx@163.com

Abstract. In recent years, the use of computer technology in medical image processing and analysis has been a hot issue of research and application at home and abroad, and the application of the PACS system has created a new era of digital medical treatment. Based on the technology of PACS system for post-processing segmentation, an improved watershed algorithm has been brought up to the medical image segmentation. Compared with the original algorithm, the improved algorithm greatly improves the accuracy of image segmentation; and the problem of over-segmentation of the original algorithm has been solved better. Dynamic merge rule is added in process of implementing watershed algorithm and the formation of over-segmentation regions due to the algorithm itself are restrained directly. Applying the improved algorithm to the PACS system, it has been proved that the algorithm can effectively extract the medical image lesions.

Keywords: PACS system, Medical image segmentation, Watershed algorithm, Dynamic merge.

1 Introduction

Nowadays, the combination of computer technology and medical imaging technology has been the main trend. This combination takes good news to both doctor and patient, because it can raise the accuracy of clinical diagnosis. PACS, Picture Archiving and Communication System, is the typical representative of hospital informatization. PACS is a comprehensive system that can collect, store, manage, diagnose and post process the images generated by kinds of digital medical equipments [1].

As a platform of digital application of medical images, PACS can transfer images that generated by various equipments to one format. Medical Image Segmentation of PACS System Image Post-processing is studied in this paper. Image segmentation is a way by which we can separate the area from the surrounding. Such segmentation can help the doctor to get the specify part from the whole picture so as to eliminate interference from irrelevant parts. This is very essential for doctor to make correct judgment about the disease and decide the correct diagnostic strategy.

*

Fund: Supported by Guangxi, Yulin Normal University youth research projects in 2010(No:2010YJQN28).

(2)

Due to characteristic of medical imaging technology, compared with the normal picture, medical image is not so contrasting to background, unclear in edge, with more noise, ect. In recent years, Watershed Algorithm [2] is of great concern due to its effectiveness in operation and precise edge locating. However, serious over-segmentations generally exist. How to overcome such disadvantage has been a major concern in the research. According to the feature of rich detail information in the medical images, many methods are comprehensively used to get the segmentation results before and during the segmentation. First, pre-process the noise in a mathematical morphology method before using Watershed Algorithm. Second, impose restrictions directly on the formation of watershed. The detailed segmentation process is showed in Figure 1.

Fig. 1. Segmentation process

2 Mathematical Morphology

Mathematical Morphology is a method applied in image process and pattern recognition. Its basic idea is to measure and extract the shape via some morphological structuring elements, so as to analyze and recognize the image. Mathematical Morphology’s basic operation consists of dilation and erosion, and the opening and closing operation based on the extension of the two basic operations above.

Erosion and dilation has most impact in the image edge, so we can get image edge by operation between the treated image and original one. However, erosion and dilation has little filtering effect of noise. Morphological opening operation and Morphological closing operation are obvious more effective at filtering. But since it take use of the complementary between erosion and dilation, the result just reflect image edge concavity and convex [3], but not all the features of the image edge. The image segment method based on the Watershed Algorithm depends largely on the algorithm we use to get the gradient. In view of this, a extended mathematical morphology method to get the image gradient and pre-process the image are suggested in this paper. This algorithm takes use of following operation.

Original image

Pretreatment(Extended

mathematical morphology filtering)

Output

Watershed segmentation (using the Dynamic Merge Rules)

(3)

(M

o

B)⊕ B-M

B

(M=(F

o

B)

B, B is 3×3 square structure element. F is the original image.) First of all, pre-process the image by morphological opening operation and morphological closing operation so as to filter the noise. Second, use Morphological opening operation to smooth the image .Third, dilate the image. The difference between the original and the one get by dilation is the better image edge.

3 Dynamic Merge Rules

By the step introduced before, we can get the target edge and gradient image reflecting all the edge. Ignore the minute irrelevant object or the details inside. And apply the gradient image processed by noise filter as input image for Watershed Algorithm. However, for the image segmentation based on Watershed Algorithm, minute changes in some factor, such as background, noise, object edge and inside, will result in the division of the same object[4]. In other words, they will be affected by noise and error. Any minimum point will be treated as a valley, and any valley will become a basin while segmented by Watershed Algorithm. Then problems come, the uniform region in image will be divided into many subregions. This makes the result disorganized, and brings many fake edges. This means failure in segmentation. This is what we call over-segmentation.

In this paper, we bring Dynamic Merge rules into Watershed Algorithm, so as to restrain the image divided into subregions unnecessarily, thus eliminating over-segmentation and achieve better effects.

(1) Rearrangement of the gradient graph

As introduced before, we can get the original image gradient graph by extended mathematical morphology operators. The gradient image pixels ordered by spatial relations, but Watershed Algorithm can not process the image in such order. Therefore, we must rearrange the gradient image to meet the algorithm requirement. Watershed Algorithm processes the image from minimum pixel gradient to maximum. Therefore, all the pixels in the gradient map are arranged according to their pixel values in ascending order. The implementation of the algorithm requires repeated access to the neighborhood of each pixel. In order to improve computing speed, we manage to store 8 neighborhood pixel information instead of 4 for each pixel during rearrangement.

(2) Segmentation and merge

After gradient graph rearrangement, algorithm dealing with the image from the pixel with minimum gradient to the one with maximum gradient, this is the entire segmentation and merger process.

First, we must consider which storage solutions we should take for all the pixels to be split. The segmentation is not to process all the pixels in one time, so we don’t need to scan the whole image each time. Because of this, we apply First In First Out (FIFO), one kind of circular queue in data structure, in the program to complete the process. This storage structure is not only helpful in storing and finding the pixel quickly, but also helpful in speed popping up the pixel. So the whole segmentation process speeds up.

(4)

Based on the algorithm process order, if the pixel with gradient value h disposed, it means that the pixels of gradient value less than or equal to h are all disposed. All the disposed pixels can be treated as a set denoted as M. Then we consider about the pixels with gradient h+1, which are not disposed. We denote it as set N. Then set M and set N may have the following three relationships.

a、N and M have no value shared by each other. This means all the pixels around set N has have gradient values higher than or equal to h+1. So N is the new valley.

b、N and M have value shared by each other, and they connected. N belongs to the valley of M.

c、N and M have value shared by each other, however, they not connected. This means N belong to the other basin [5]. By calculating the distance between the bottom and N, we can find where N should locate.

The following introduce the concrete steps. It is divided into four steps:

① Put the pixels into the queue. First, put the pixels into circular queue under First In First Out (FIFO) rule, then output it at the appropriate time. In other words, if segmentation process request to process pixel with gradient value h, the circular queue response for put the right pixels into the queue. When the process finish, the circular queue should release the disposed one.

② Pixels out of queue. When the input finished, circular queue is responsible for releasing the disposed pixel. We denote the pixel to be released as P. We calculate the basin and the watershed in neighborhood of pixel P, and denote them as NumB and NumW. The affect NumB and NumW have on output of the watershed will be discussed following.

③Judge the output for P based on the NumB and NumW.

If NumB = 1, it means that only one basin adjacent to P. So we should assign the output to P.

If NumB = 0 and NumW> 0, it means that the watershed is P.

If NumB > 1, it means that there are many basins in P’s adjacent area. According to the original algorithm, P is the watershed. In this case, there will be over-segmentation. Now we bring the Dynamic Combination Rule to Watershed Algorithm to solve this problem. First, find out the merging mother town for point P, and denote it as W. The principle is to find the point in the adjacent basins. The one in the deepest bottom is W.As for the other basins, calculate the dynamic value for the bottom. Appoint a threshold value first. Then compare the bottom dynamic value with this threshold value. If less than the threshold value, all the pixels of this basin should be incorporated into mother town W of P. Repeat this operation. If all the basins merged into W, output of W will be assigned to P. Otherwise, P is watershed.

④ Judge whether there is a new bottom

If there is no pixel with value lower than h+1, pixel with value h+1 should correspond the bottom of the new basin. At this moment, the algorithm should scan the whole image to judge whether there is a new bottom.

(5)

4 Image Segmentation and Analysis

Select tuberculosis CT image from PACS for segmentation, results are shown in Figure 2.

.

Fig. 2. Segmentation results comparison of Tuberculosis CT images before and after using the improved watershed algorithm

Figure 2a is a typical CT image of tuberculosis, the small white dot in the left and right lungs is the tuberculosis. After enlarging and stretching the image, doctor can not make the accurate diagnosis. In order to analyze the image better, we must have the segmentation in the image. After detecting the edge by extended mathematical morphology operators, we get result as Figure 2b. If we do segmentation with traditional Watershed Algorithm, we get result as Figure 2c. You can see that the segmented image makes no sense, because it’s over-segmentation. Have segmentation on Figure 2b by improved Watershed Algorithm, we get result as Figure 2d. We can see that segmentation result is very clear, lesion area have been separated from the surrounding successfully. This allows doctor to focus on analyzing lesion area, so doctors can make further diagnosis to the patient.

All operations are running in PACS system in computer with 2.5GHz Intel Pentium IV processor and 1.99GB RAM.

b. Extended operator graph a. Original tuberculosis

c. Segmentation results before Improved

d. Improved segmentation results

(6)

5 Summary

This paper introduces an improved Watershed Algorithm to solve the problem of over-segmentation. First, an extended mathematical morphology operator is proposed to process the original image. And then, Dynamic Merge Rule Is added to Watershed Algorithm to improve it. Applying the improved Watershed Algorithm to the PACS system, we can have segmentation in the image; moreover, extract the lesion area from whole image. The result shows that the improved algorithm has good noise immunity and works well at medical image segmentation. Therefore, it can help to improve the diagnosis and treatment.

References

1. Fang, Q.-L.: Biomedical Engineering Shanghai 1, 58 (2003)

2. Vincent, L., Soille, P.: IEEE Trans. on Pattern Analysis and Machine Intelligence 13, 583 (1991)

3. Zhao, Y.-Q., Gui, W.-H., Chen, Z.-C., Li, L.-Y.: Computer Engineering and Applications 33, 6 (2005)

4. Lui, H.-T., Zhang, Y.-J.: Data Acquisition and Processing 12, 18 (1997) 5. Bieniek, A., Moga, A.: Pattern Recognition 21, 907 (2000)

References

Related documents

Concerning transitivity and the activity in which social actors are involved, as Table 4 indicates females are mostly activated in relation to material

Figure 12-2 below shows the soil plasticity index for soils in the City of Plano and the areas most susceptible to problems with expansive soils. City of Plano Expansive and

Our findings suggest that although it can be argued that Chinese SOEs are subject to more government interferences and Chinese private firms are subject to more market

LETTER Earth Planets Space, 63, 853?857, 2011 The resonant response of the ionosphere imaged after the 2011 off the Pacific coast of Tohoku Earthquake Lucie M Rolland1, Philippe

Participants ascribed the target less agency and expressed greater empathy towards him when his obesity was attributed to a psychological compared behavioural aetiology, and

recoverable CFU and a 15% increase in bio fi lm removal ef fi cacy compared to treatments that had only Am or Ag. More importantly, PDA-assisted treatment was found to immobilize Ag

DELAYED EXPRESSION OF MUTATIONS 455 Where the mutation rate per bacterium per division cycle is a, and the rate per bacterial division is a/(ln 2), the proportion of