Echogenicity measuring of ultrasound images based on multi-agent system to appraisal of diagnosis

Transcription

1 Echogenicity measuring of ultrasound images based on multi-agent system to appraisal of diagnosis JIŘÍ BLAHUTA, TOMÁŠ SOUKUP, PETR ČERMÁK, DALIBOR HULA Department of Informatics Silesian University in Opava Bezručovo nám. 13, Opava CZECH REPUBLIC MICHAL VEČEREK Faculty of Electrical Engineering and Computer Science VŠB-TUO Ostrava Tř. 17. listopadu 15, Ostrava CZECH REPUBLIC Abstract: - This presented paper is focused on image processing of transcranial sonography (TCS) images to Parkinson s disease recognition by means of own developed MATLAB application, more precisely on the first phase of pre-processing of TCS images. This phase is realized by using multi-agent (MAS) system which analyze brightness inside ROI substantia nigra, more precisely echogenicity substantia nigra to potential risk of parkinsonism. Firstly, we detect global brightness of cut with ROI and subsequently only locally inside substantia nigra to detect lesions. Furthermore, paper also briefly presents phase of the main processing with exact measuring SN area and post-processing with saving the measured values to further analysis. Key-Words: - MATLAB, ultrasound, substantia nigra, agent, Parkinson, percentile, blocks, submatrix, diagnosis 1 Introduction Medical imaging is very important in modern medicine. The main benefit is using of DICOM format for all used modalities; such as ultrasound, CT, etc. We developed a MATLAB-based application with GUI for processing of ultrasound TCS images to potential Parkinson s disease (PD) detection. This application is focused on ROI-based processing with artificial intelligence elements, more precisely using of a multi-agent system. This application has been tested for different US devices and artificial intelligence elements which help to semi-automatic detection. In the text below we will present the first phase of the processing; using of MAS to local echogenicity detection. This phase is useful as the first warning that are detected suspicious lesions in ROI which may means PD. This warning can be confirmed in the main phase of the processing where we measure SN area depend on binary threshold, see below. 2 Loading an image and cut Generally, this application is useful to potential detection of PD which is generally characterized by damaging substantia nigra (SN) area. SN is an elongated nucleus situated in each cerebral peduncle lateral to the red nucleus with cells containing melanin 1 and produce dopamine to correct function of the CNS. The main criterion is the area of SN, risk threshold is cm 2. So, we will measure risk threshold area to detection PD and non-pd cases. Critical analysis of PD is described in details in [14]. PD is caused by the death of dopaminergic neurons. It is a degenerative disease of basal ganglias inside the brain. The main symptoms of PD include muscle rigidity, tremor and changes in 1 bstantia_nigra.aspx?s=substantia+nigra&mode=1&syn=&scope = This research has been supported by grants SGS 6/2011. ISBN:

2 speech and gait, bradykinesia, sleep disorders and more. 2 Detailed information about SN and PD symptoms and US background are available in [5], [7], [9] and [11]. The following figure shows the brain-stem area 3 and corresponding TCS image with mm window to cut. Furthermore, application automatically converts RGB input into grayscale by means of the following equation (1). (2) So, we consider an image in form of the discrete matrix DI as follows:, (3) and each pixel p is represented as p(x, y) up to p(m, n) with gray level intensity I. It is necessary to further processing by MAS and also to measuring of the area. The following flowchart introduces the main steps of our designed processing. 3 Using of multi-agent system to lesions detection Multi-agent systems (MAS) 5 are strong tool of distributed artificial intelligence for different branches which include image processing. MAS are interdisciplinary science based on cognitive science robotics and related disciplines of artificial intelligence. Generally we can describe MAS as a finite set of n autonomous agents:, (4) Fig. 1 Substantia nigra in the brain and corresponding TCS image To further processing we will need 50x50 mm area only which is highlighted on Fig. 1. Application also provides automatically cut the input image depend on device preset set of images from the same device has the same resolution and we can simply cut mm area from axis which is computed from Euclidean distance for two points with coordinates (x 1, y 1 ) and (x 2, x 2 ) in case of 2-D space, therefore in 2-D discrete image is given by the Pythagorean theorem 4 as follows: (1) which can communicate and cooperate and are situated in appropriate environment, in our case agents are situated in 2-D image, more precisely image matrix. Each agent is autonomous and can communicate with other agents within MAS and solve the problem together. In our research we use software agents only, thus MAS is represented as an autonomous source code. More information about MAS and using are available in [8]. After the initial processing including load of the input image and cut, we need the second part which is applicable for MAS. We can use MAS to checking intensity in two steps, more precisely with using of 2 agents which communicate. In accordance with the fact that B-MODE 6 is static ultrasound imaging which is primarily based on gray-level intensity according to tissue density, /b_mode.aspx?s=b-mode&mode=1&syn=&scope= ISBN:

3 checking of image quality and subsequently detection of lesions in SN is based on intensity of gray. The brightness of each dot is determined by the echo amplitude. The following flow-chart describes using of the agents in this instance: Each submatrix DI i contains finite number of pixels depend on image resolution. For our work are not decisive the number of pixels, we need to work with real dimensions from native US axis in DICOM image. Each submatrix (block) is represented by 1 intensity value computed from intensities of pixels inside the block. Fig. 3 Block processing of the input image 50x50 mm to image quality detection Each block DI i for i=1 to n (in this instance n=64) (submatrix) is represent as average value of the pixels inside the block. We consider the submatrix DI i which contains the pixels with coordinates p(x,y) as follows:. (6) Indexes u, v represent resolution of the block DI i, thus number of pixels in the block. For each DI i was computed average value of intensity which represents the block:. (7) Fig. 2 Flow chart of MAS processing 3.1 Agent to global analysis We will work with image mm as we described in the previous chapter. We divide input image, thus matrix DI, into 8 8 subregions. Each block represents the average intensity value of pixels which are inside block, thus we get 64 values. Mathematically, we divided image matrix DI into 8 submatrices DI 1, DI 2,,DI 64 (8 8) as follows:. (5) So, agent has environment represented by 2-D block from image. Inside each region (block) we constructed independently an agent which must detect if inside region average intensity exceeds 25, moreover is computed minimal value of intensity. This value has been MAS is composed by these agents which operate inside each block independently. The following figure shows an original input mm and divided into blocks for agent detection. Subsequently is determined block with minimal intensity from the blocks (5): (8) where DI i is i-th block (submatrix DI i ). ISBN:

4 In other words, we must check that minimal intensity from blocks does not exceed 25. If this minimal intensity exceeds 25, input image may be corrupted by manual set of intensity and lesions may be judged as pathology in spite of that patient is physiological (non-pd). 3.2 Agent to local SN analysis Agent which detects global intensity of the input image and its average in blocks, is useful to global properties of image. In other words, this agent detect that input image is not corrupted by over-brightness. Agent to local SN analysis is similar to previously described but now agent will operate only inside SN ROI. In other words, agent will not detect intensity globally from whole image mm but only for SN area. Firstly, with regard to smaller area of SN than the whole image mm, we must change division into blocks from 8x8 into blocks (submatrices), thus we get 1024 blocks. In MATLAB is given by blockproc function: fun mean2(block_struct.data)* ones(size(block_struct.data)); relg2 = blockproc(greyimage,[d/32 d/32],fun); Fig. 4 Blocks 32x32 to local analysis of the average intensity values in ROI SN only Where x is x-resolution of image and y represents y- resolution. From variables xc and yc which represent center c(xc,yc) we measured where is SN located in consideration of c. Hence, SN area was computed: wsnl = imcrop(relg2,[xc-40 yc ]); In case of using of different set of images will be variables xc and yc different in consideration of image resolution; different mm window. In the previous case of the global image analysis we found block with minimal intensity value (8). Computing of minimal value of intensity is nonobjective regard to SN area. In the previous case we computed global brightness of image to determination if an input image is corrupted or not. In case of local SN intensity we will compute average value from all 1024 blocks where each of them is represented by average value (6) of pixel intensity, see the previous chapter. We distinguished 3 cases of detected average value to suspicion on PD: if ag2 < 20 msgbox(int2str(ag2) 'is the average intensity value in SN, No lesions with high gray-level intensity, there is no risk of PD') elseif 20 < ag2 < 25 msgbox(int2str(ag2) 'is the average intensity value in SN => potential risk of PD') else > 30 msgbox('is the average intensity value in SN => patient with risk of PD') end and subsequently we will get one from the message about average value inside ROI SN: The agent work only in presented cut of SN area area; will be eliminated neighboring blocks which may be undesirable to local brightness level. Thus, this agent computes average value (3) inside SN area to potential occurrence of lesions inside SN. Due to equal resolution of used set of images we know where is SN located. We computed the center c of the image with coordinates xc and yc: (9) Fig. 5 Dialog with the average gray-level intensity and risk of PD For different images, where the global gray-level intensity, agent may be modified for a new set of images and rules to distinction of PD and non-pd ISBN:

5 may have different values of the gray-level intensity. The algorithm is based on the following steps: create image divided into 32x32 blocks; each of them represents average intensity of pixels agent go through the blocks and checks average intensity in ROI SN the result of potential PD is given by the previous IF-ELSEIF-ELSE condition depend on average intensity value This method to find the average value of graylevel intensity efficiently eliminates isolated pixels (4 or less). It is very important to the correct detection. If in SN will be an isolated pixel with high intensity (T > 30) and another will be very dark down to T = 0, the average value does not exceed 30 and this isolated pixel will not be considered as the lesion to PD. If the first agent, which detect minimal intensity value on image divided into blocks 8 8, detect minimal intensity exceeds T = 25, the second agent will not continue for local intensity detection in SN. The first agent hands information about global intensity to the second agent. 4 ROI-based main processing To further processing after MAS-based detection we need ROI-based processing. We require the elliptical ROI with area A = 50 mm 2 with rake angle of 60, which is needed to ROI definition. Shape, size and rake angle of ROI were assigned by neurologist. The following figure shows the example of used ROI. block processing 1 1 mm. The area has been computed from binary image number of blocks to get the real area in mm 2. The main goal of this ROI processing is computing of area inside ROI to detection pathology or physiology which is depending on black or light regions inside ROI and we will get the following graph with 90 th percentile curve which has been computed from a set of non-pd patients. Fig. 7 Measuring of the SN area in mm 2 depending on the intensity T with referential curve of 90 th percentile Shape of these curves depends on decreasing area in ROI depending on intensity. Firstly, from a set of 100 images, we computed 90 th percentile from ROI-based area measurement as we described. This curve of 90 th percentile represents the border between non-pd and PD patients. Subsequently is computed difference between 90 th percentile as border and curve which represents measuring of examined patient (10). This difference was computed by sum of measuring values of ROI area (256 values for intensity T = 0,1,2, 255) for an examined patient and referential curve of 90 th percentile. The referential curve was computed by once and is fixed., (10) where ap is the ROI SN area of the examined patient and ar is the ROI SN area for the referential 90 th percentile curve. More information about reproducibility of the used method are available in [3] and [4]. Fig. 6 Used ROI to the main processing This phase is based on binary thresholding in ROI area and computing of area of defects in this ROI for all intensity levels T 0; 255. We will get a graph with computed area which is computed from 5 Conclusion In the presented paper has been showed and described using of MAS to medical image processing. This MAS has been designated for preprocessing phase of ultrasound TCS images. The MAS has two main goals. Firstly, checking of image quality on input based on block-processing with 8 8 blocks (grid) with average values of ISBN:

6 intensity and subsequently finding of minimal intensity in the image. If minimal intensity exceeds 25, the input image might be corrupted. The second part of the MAS is used to finding of average value of intensity only in ROI SN area to potential PD diagnosis. With regard to smaller area was designed the grid with smaller blocks 32x32. This agent which checks average value only in ROI SN communicates with the first agent which checks global quality of the input TCS image. If the first agent detects that input image may be corrupted due to badly setup of ultrasound gain, the second does not work; in ROI SN will be find incorrect values due to incorrect input. The agents are designed as isolated part of the source code in our developed MATLAB application. These agents work before the main processing and secure a fast basic image analysis. The second agent has 3 different states which determine that patient is probably diseased or not. After this phase pre-processing based on MAS for testing of the image quality and local average intensity follows the main processing based on ROIbased computing of the area in mm 2 in ROI SN depending on the intensity T and subsequently is appraised that examined patient has PD features (lesions in SN) or not by the referential curve as border of non-pd patients. The future work will be focused on improving of MAS. Agents should be adaptable with learning of the gray-level intensity by re-inforcement learning method from different sets of ultrasound images. Thus, agents would learn to recognize different gray-level intensity of the lesions depending on global average level of intensity, resolution and position of SN area. The detection of the lesions in SN will be more accurate and intelligent depending on input image with different global gray-level intensity. Moreover, number of the agents should be higher and each of them should find different criterion such as number of isolated pixels, relative gray-level intensity according to the image, etc. References: [1] Becker, G.: Degeneration of substantia nigra in chronic Parkinson s disease visualized by transcranial color-coded real-time sonography, 1995, Journal of Neuroimaging 45. [2] Blahuta J., Soukup T., Čermák P., The image recognition of brain-stem ultrasound images with using a neural network based on PCA, Medical Measurements and Applications Proceedings (MeMeA), 2011 IEEE International Workshop on BARI, 2011, ISBN: [3] Blahuta J., Soukup T., Čermák P., The ROI defect statistical analysis of substantia nigra to reproducibility of designed experimental algorithm for potential PD diagnosis, Mathematical Methods and Techniques in Engineering, Catania,Wseas Press, ISBN: [4] Blahuta J., Soukup T., Čermák P., Automatic ROI positioning in ultrasound TCS image using artificial intelligence to Parkinson s disease risk, Recent Researches in Applied Information Science, 2012, Faro, Wseas Press, ISBN: , pp [5] Gelb, D., O. E. G.-S.: Diagnostic criteria for Parkinson s dinase, Archives of Neurology 56(1), 1999, pp [6] Principe J. C., Euliano N. R., Lefebvre W. C., Neural and Adaptive Systems: Fundamentals Through Simulations, 2000, John Wiley & Sons, Inc. [7] Schreiber, J., Sojka, E., Ličev, L., Škňouřilová, P., Gaura, J., Školoudík, D.: A new method for the detection of brain stem in transcranial ultrasound images, Proceedings of Biosignals 2008, [8] Shoham Y., Leiton-Brown K.: Multiagent Systems: Algorithmic, Game-Theoretic, and Logical Foundations, Cambridge University Press, 2008, ISBN [9] Školoudík, D.: Reproducibility of sonographic measurement of the substantia nigra, Ultrasound in Medicine & Biology (9), 2007, pp [10] Walter U, Wittstock M, Benecke R, Dressler D: Substantia nigra echogenicity is normal in nonextrapyramidal cerebral disorders but increased in Parkinson's dinase, J Neural Transm. 2002, 109:, pp [11] Webster DD: Critical analysis of the disability in Parkinson's disease, Mod Treat, 1968, 5:, pp ISBN: