Automated Classification Of Ultrasound Kidney Images Computer Science Essay

Published:

This essay has been submitted by a student. This is not an example of the work written by our professional essay writers.

Ultrasound imaging is a technique of viewing the internal organs of the body. The principle involves exposing that part of the body to high frequency sound waves. They show the organs' structure and movement, as well as blood flowing through blood vessels. In the Kidney there are various abnormalities. Of these, Cyst and Medical renal diseases are most commonly found. The medical images are preprocessed using various filters and their performance is compared. The Region Of Interest (ROI) is Segmented using region marker and semi-automated methods. The features considered here are statistical and textural parameters. First method is done using GLCM Algorithm. Features are extracted using this method at different angles and orientations. Out of the extracted features, 14 were found to be highly useful for the classification. The various features that are used are autocorrelation, contrast, correlation, cluster Prominence, homogeneity, energy, entropy, difference energy, difference entropy, mean, standard deviation, skewness, and kurtosis. The obtained features are given to the neural network classifier. The main aim is to identify whether the given image is normal or abnormal (presence of Cyst or Medical Renal disease). All these parameters from the various modules are fed into Statistical Classifier (Naïve Bayes Classifier) and also to 3 types of Neural Network classifiers- MLP (Multilayer Perceptron), SVM (Support Vector Machine) and PNN (Probabilistic Neural Network) and the classification are done.

CHAPTER - 1

INTRODUCTION

PURPOSE OF THE PROJECT

Ultrasound imaging is the most widely used method of imaging human abdominal

Organs like kidney, spleen and liver. Ultrasonic echoes from human tissues displayed as a B-scan image form a texture pattern. One application of diagnostic ultrasound is kidney imaging. Kidney being a vital organ for existence, identifying kidney abnormalities is of prime importance. Because these abnormalities are concentrated in a small area of the kidney tissue, while the rest being normal, finding the abnormality sometimes becomes tough. Also the ultrasound images are degraded by speckle noise and other artifacts thereby making it difficult for the extraction of cysts and discriminating MRD from normal.

The kidney is the part of the urinary system. It plays a major role. There are a few abnormalities present in the kidneys such as Cysts and Medical renal disease.

The renal cyst is a benign non-cancerous mass developing in renal parenchyma (cortical cyst) or within the sinus region (parapelvic cyst). The cyst cavity is unilocular, filled with clear serous fluid. It is also lined with a cuboidal epithelium but without any communication to the renal collecting system. These cysts arise because of tubular or lymphatic obstruction.

Medical renal diseases involve mainly the parenchyma of the kidneys. Hematuria, proteinuria, pyuria, oliguria, polyuria, pain, renal insufficiency with anaemia, electrolyte abnormalities, and hypertension may occur in a wide variety of disorders affecting any portion of the parenchyma of the kidney, the blood vessels, or the excretory tract. [1]

The aim of this project is to detect these abnormalities and classify them into their respective category.

OBJECTIVE OF THE PROJECT

The objective of the project is to compare the classification techniques for identifying kidney abnormalities. The desired region of the kidney is segmented. The main concept of texture analysis involves feature extraction. Once the feature vectors are obtained then the Bayes Classifier and Neural Network is applied to classify. The different techniques are compared to evaluate the best performance of classification.

CHAPTER - 2

LITERATURE SURVEY

PAPER TITLE

TECHNIQUE APPLIED

RESULTS

LIMITATIONS

OUTCOMES

ADVANTAGES

1.Sonography Images for Breast Cancer Texture classification in Diagnosis of Malignant or Benign Tumors

Textural features estimation using GLCM algorithm.

1.For the ultrasound images promising results were obtained by using GLCM in Different angles and different pixel distances. 2. It produced distinct variations in the textural features.

Care has to be taken to estimate the important textural features with minimum complexity.

1. Distinct features were obtained in all pixel distances. 2. The textural features are important in the case of ultrasound images 3. Excellent tool that can be used for ultrasound feature extraction.

In this paper GLCM has been carried out in 4 pixel distances and 4 angles. This helps in tracing the textural features in various subsections of the image.

2. Analysis of Ultrasound kidney Images using Content Descriptive Multiple Features for Disorder Identification and ANN based Classification

Features are extracted using various algorithms and this is used as input to the MBNN Classifier. Training data-150 images Testing data-50 images

1.Various features that are of high importance are extracted 2. The MBNN Classifier has provided around 85% efficiency.

1. The Computational complexity is slightly high. 2.Many features that are of less importance and not used for further analysis has been extracted

1. Kidney images were extracted using spline interpolation method. 2.1st order and 2nd order statistical features were extracted. 3.The MBNN Classifier was used, which yielded an efficiency of around 85%

One of the best methods to do the processing of Ultrasound Kidney images. It helps in the automated classification of kidney images into 3 classes.

3. Breast Tissue Classification using Statistical Feature Extraction of Mammogram.

The statistical features extracted are the mean, standard deviation, smoothness, third moment, uniformity and entropy which signify the important texture features of breast tissue. Based on the difference in values they classify it.

1.Classification of normal and cancerous breast. 2. Early detection of breast cancer.

1. The selection of ROI for the application of the algorithm. This process is very difficult in this technique.

Experimental results show a visual clue to any radiologist about the abnormality in the growth of breast.

Signifies the important texture features of breast tissue. Based on the values of these features of a digital mammogram, the authors have made an attempt to classify the breast tissue into various basic categories.

4. Removal of Speckle Noise from Ultrasound Medical Image based on Special Filters: Comparative Study.

Various filters are used to remove the speckle noise from ultrasound prostate images.

1. For the ultrasound images promising results were obtained by using Median and M3 filter. 2. It produced distinct filtered image in which the boundaries are seen. 3. Smoothing resulted in suppressing the variations within the same texture region in the output image, thus enhancing the segmentation process significantly.

Most of the filters failed to completely remove multiplicative speckle noise.

Filtering method to remove the multiplicative speckle noise and comparison of statistical parameters like PSNR , SNR to estimate the better filter.

The quality of the medical ultrasound images have improved and the medical images are now ready to be used for further analysis.

5. Localizing Region-Based Active Contours

Segmentation of the ROI using Active contours method.

1. Improved accuracy in segmenting the image of interest. 2. Useful particularly for heterogeneous images.

1. This particular algorithm usually settles with local minimum. 2.Takes lot of iterations to settle in the desired ROI

The experimental results show that this method is best suited for heterogeneous objects like ultrasound kidney images.

This method shows an easy to understand and quite accurate method to extract the ROI.

CHAPTER 3

ULTRASOUND IMAGING

3.1 INTRODUCTION

Ultrasound imaging, also called ultrasound scanning / sonography, involves exposing the organs of the body to high-frequency sound waves to produce pictures of the inside of the body. Ultrasound imaging avoids using ionizing radiation and therefore it is a noninvasive medical examination that helps doctors diagnoses various medical conditions.

Ultrasound imaging is done in real-time and they can show the structure and movement of the various internal organs, and also blood flowing through blood vessels.

Ultrasound technology has undergone tremendous improvements over the years. Three-dimensional (3-D) ultrasound is one such recent advancement. It transforms the sound wave into 3-D images. Four-dimensional (4-D) ultrasound is 3-D ultrasound in motion.

A Doppler ultrasound study is also now being used as a part of an ultrasound examination. It is a special ultrasound technique that evaluates blood flow through a blood vessel, including the body's major arteries and veins in the abdomen, arms, legs and neck.

3.2 PRINICPLES OF ULTRASOUND IMAGING

Ultrasound imaging is based on the same principles involved in the sonar. When a sound wave strikes an object, it bounces back, or echoes. By measuring these echo waves it is possible to determine how far away the object is and its size, shape, and consistency (whether the object is solid, filled with fluid or both).

In an ultrasound examination, a transducer both sends the sound waves and records the echoing waves. When the transducer is pressed against the skin, it directs small pulses of inaudible, high-frequency sound waves into the body. As the sound waves bounce off of internal organs, fluids and tissues, the sensitive microphone in the transducer records tiny changes in the sound's pitch and direction. These signature waves are instantly measured and displayed by a computer, which in turn creates a real-time picture on the monitor. One or more frames of the moving pictures are typically captured as still images.

3.3 COMMON USES OF THE PROCEDURE

Ultrasound examinations can help to diagnose a variety of conditions and to assess organ damage following illness. Ultrasound is used to help physicians evaluate symptoms such as:

Pain

Swelling

Infection

Ultrasound is a useful way of examining many of the body's internal organs, including, but not limited to the:

Heart and blood vessels, including the abdominal aorta and its major branches

Liver

Gallbladder

Spleen

Pancreas

Kidneys

Bladder

Uterus, ovaries, and unborn child (fetus) in pregnant patients

Eyes

Thyroid and parathyroid glands

Scrotum (testicles)

Ultrasound is also used to:

Guide procedures such as needle biopsies, in which needles are used to extract sample cells from an abnormal area for laboratory testing.

Image the breasts and to guide biopsy of breast cancer.

Diagnose a variety of heart conditions and to assess damage after a heart attack or diagnose for valvular heart disease.

BENEFITS OF ULTRASOUND

Ultrasound scanning is noninvasive (no needles or injections, in most cases) and is usually painless.

Ultrasound is widely available and easy to use.

Ultrasound imaging uses no ionizing radiation, and is the preferred image modality for diagnosis and monitoring of pregnant women and their unborn infants.

Ultrasound provides real-time imaging, making it a good tool for guiding minimally invasive procedures such as needle biopsies.

Ultrasound images can visualize structure, movement and live function in the body's organs and blood vessels.

3.5 PROBLEMS IN ULTRASOUND IMAGING

The echoes are displayed that do not correspond to real structures.

Sometimes the genuine echoes are omitted from the display.

There is displacement of echoes.

The echo characteristics are subjected to severe distortion.

3.6 LIMITATIONS OF ULTRASOUND IMAGING

Ultrasound waves are disrupted by air or gas; therefore ultrasound is not an ideal imaging technique for the bowel or organs obscured by the bowel.

Large patients are more difficult to image by ultrasound because tissue attenuates (weakens) the sound waves as they pass deeper into the body.

Ultrasound has difficulty penetrating bone and therefore can only see the outer surface of bony structures and not what lies within.

3.7 KIDNEY ULTRASOUND IMAGING

Renal ultrasound is done when there's a concern about certain types of kidney or bladder problems. Renal ultrasound tests can show:

The size of the kidneys

Signs of injury to the kidneys

Abnormalities present since birth

The presence of blockages or kidney stones

Complications of a urinary tract infection (UTI)

Cysts or tumors

3.7.1 Scanning Technique

A 3.5-5 MHz probe is typically used to scan the kidney.

For the right kidney, the patient is made to lie supine and the probe is placed in the right lower intercostal space in the midaxillary line. The liver is made as "acoustic window" and aim the probe slightly posteriorly (toward the kidney).  Gently rock the probe (up and down or side to side) to scan the entire kidney.  If needed, the patient can be made to inspire or exhale, which allows for subtle movement of the kidney.  Obtain longitudinal (long axis) and transverse (short axis) views.

Figure1: Normal right kidney in longitudinal axis

Figure2: Normal right kidney in Transverse axis

For the left kidney the patient is made to lie supine or in the right lateral decubitus position. The probe is placed in the lower intercostal space on the posterior axillary line.  The placement will be more cephalad and posterior than when visualizing the right kidney.  Again gently rock the probe to scan the entire kidney.  Obtain longitudinal and transverse views.

Figure 3: Normal left kidney in longitudinal axis

Figure 4: Normal left kidney in transverse axis

Depending on which axis is used to obtain the images, the sonographic shape of the kidney will change.  On longitudinal view, the kidney will appear football-shaped and will typically be 9-12 cm in length and 4-5 cm in width (normally within 2 cm of each other).  On transverse view, the kidney appears C-shaped.

The normal kidney will have a bright area surrounding it which is made up of Gerota's fascia and perinephric fat.  The periphery of the kidney will appear grainy gray which is made up of the renal cortex and pyramids. The central area of the kidney, the renal sinus, will appear bright (echogenic) and consists of the calyces, renal pelvis and the renal sinus fat.  Always both kidneys are scanned for comparison and correlation to clinical picture. 

CHAPTER - 4

PROCESSES INVOLVED

4.1 INTRODUCTION

The main aim of the project is to detect the type of abnormalities present in the ultrasound kidney images. The various processes involved are:

Getting the input image.

Preprocessing of the images.

Segmentation to obtain the kidney separately.

Feature extraction -statistical and textural.

Classification is done using different Neural Networks.

4.2 FLOW OF THE PROJECT

Figure 5: Flow Diagram

4.3 PREPROCESSING

The images acquired from the scanning center are preprocessed. First, the image is converted to gray scale. It is then preprocessed to remove multiplicative speckle noise using various filters. Then the contrast enhancement is carried out to make the image suitable for processing.

4.4 SEGMENTATION

The pre-processed images are segmented to obtain the Region of Interest (ROI) using semi-automatic and Region marker methods of segmentation. The whole medical information is present only in the small region of the image and this ROI once taken is used in further processing.

4.5 FEATURE EXTRACTION

Both statistical and textural parameters are considered for further processing. The textural features are extracted using GLCM Algorithm. Some statistical features like kurtosis, skewness, mean and standard deviation. These features help to differentiate the 3 classes namely Normal, Cyst and Medical Renal disease.

4.6 CLASSIFICATION ALGORITHM

The various Classification techniques can be used such as Neural Network, Fuzzy Reasoning and Pattern Recognition Methods. For this application we use Bayes Classifier (statistical classifier) and artificial neural network algorithm, MLP (Multi Layer Perceptron) and SMV (Support vector Machine) and PNN (Probabilistic Neural Network) for classification. The 3 different approaches are compared and the best performing network is identified.

CHAPTER 5

PRE -PROCESSING

5.1 INTRODUCTION

A technique in which the image is digitized and various mathematical operations are applied, in order to create an enhanced image that is more useful or pleasing to a human observer, or to perform some of the interpretation and recognition tasks usually performed by humans.

Ultrasound imaging method is suitable to diagnose. The accurate detection of region of interest in ultrasound image is crucial. Usually, the contrast in ultrasound image is very low and boundary between region of interest and background are fuzzy. And also speckle noise and weak edges make the image difficult to identify the kidney region in the ultrasound image. So the analysis of ultrasound image is more challenging one.

Data sets collected by image sensor are generally contaminated by noise. The region of interest in the image can be degraded by the impact of imperfect instrument, the problem with data acquisition process and interfering natural phenomena. Therefore the original image may not be suitable for applying image processing techniques and analysis. Thus image enhancement technique is often necessary and should be taken as the first and foremost step before image is processed and analyzed. An efficient filtering method is necessary for removing noise in the images.

5.2 SPECKLE NOISE

Data dropout noise is generally referred to as speckle noise. This noise is, in fact, caused by errors in data transmission. The corrupted pixels are either set to the maximum value, which is something like a snow in image or have single bits flipped over. This kind of noise affects the ultrasound images. Speckle noise has the characteristic of multiplicative noise. Speckle noise follows a gamma distribution:

where g = gray level

5.3 VARIOUS FILTERS

5.3.1 Max Filter

The max filter plays an important role in low level image processing and vision. It is similar to the mathematical morphological operation: dilation. The brightest pixel gray level values are identified by this filter. Hence, it is used for speckle noise removal from the ultrasound medical image. It is expressed as:

where (s,t) Sx,y

It reduces the intensity variation between adjacent pixels.

5.3.2 Min Filter

The min filter plays a significant role in image processing and vision. It is equivalent to mathematical morphological operation: erosion. It recognizes the darkest pixels gray value and retains it by performing min operation.

where (s,t) Sx,y

It removes noise better than max filter but it removed some white points around the border of the region of the interest.

5.3.3 Midpoint Filter

The midpoint filter simply computes the midpoint between the maximum and minimum values in the area encompassed by the filter. It combines order statistics and averaging. The Midpoint filter blurs the image by replacing each pixel with the average of the highest pixel and the lowest pixel within the specified window size.

5.3.4 Median Filter

The median filter is a nonlinear digital filtering technique, often used to remove noise. Median filter under certain conditions preserves edges while removing noise and therefore it is the widely used one. The idea of the median filter is to run through the image entry by entry, replacing each entry with the median of neighboring entries. For 1D signal, the window is the first few preceding and following entries, whereas for 2D signals such as images, more complex window patterns are possible (such as "box" or "cross" patterns).

5.3.5 Arithmetic Mean Filter

An arithmetic mean filter operation on an image removes short tailed noise from the image at the cost of blurring the image. The arithmetic mean filter is defined as the average of all pixels within a local region of an image.

The arithmetic mean is defined as:

Pixels that are included in the averaging operation are specified by a mask. The larger the filtering mask becomes the more predominant the blurring becomes and less high spatial frequency detail that remains in the image.

5.4 EXPERIMENTAL ANALYSIS AND RESULTS

The statistical measurement could be used to measure enhancement of the image. The Peak Signal-to-Noise Ratio (PSNR) is used to evaluate the enhancement performance. If the value of SNR is larger, then the enhancement approach is better.

FILTER

PSNR

Arithmetic Mean

24.5980

Mid-point

22.9870

Max

22.6066

Min

22.353

Median

25.6796

Table 1: Performance Analysis of 5 different filters

Figure 6: Performance chart of 5 different filters

5.4.1 Output of Pre-processing

Figure 7: Input image Figure 8: Filtered image

CHAPTER -6

SEGMENTATION

6.1 INTRODUCTION

Segmentation refers to the process of partitioning a digital image into multiple segments. The goal of segmentation is to simplify and/or change the representation of an image into something that is more meaningful and easier to analyze. Image segmentation is typically used to locate objects and boundaries in images.

The result of image segmentation is a set of segments that collectively cover the entire image, or a set of contours extracted from the image Each of the pixels in a region are similar with respect to some characteristic or computed property, such as color, intensity, or texture.

6.2 DIFFERENT APPROACHES FOR SEGMENTATION

Two different segmentation methods used for contouring the kidney region is considered. Any segmentation scheme should not only contour the kidney region but must involve less complexity in achieving it. Also the scheme must be reliable for segmentation of different categories of kidney. The 2 schemes considered here are

Region-Marker Method and

Localizing Region-Based Active Contours

Both these methods are suitable for application of various categories of kidneys.

6.3 REGION MARKER METHOD (RMM)

The RMM is obtained for each image by allowing the users to draw the manual outline through visual inspection over the identified boundary by moving the cursor using mouse. Once the outline is drawn, the binary mask of the drawn outline is created. The mask is placed on the original image to differentiate it from the entire image. Then the outside region is masked and finally the region of interest is obtained, it is changed to the greyscale and then is resized to the desired size.

6.3.1 Output of RMM

Figure 9: Segmentation using Region Marker Method

6.4 LOCALIZING REGION-BASED ACTIVE CONTOURS

Active contour methods have been used widely in recent years, and have found applications in a wide range of problems including visual tracking and image segmentation. The basic idea is to allow a contour to deform so as to minimize a given energy function in order to produce the desired segmentation.Better suited for medical imagery where Heterogeneous objects is common.

Two main categories exist for active contours: edge-based and region- based. Edge-based active contour models utilize image gradients in order to identify object boundaries. This type of highly localized image information is adequate in some situations, but has been found to be very sensitive to image noise and highly dependent on initial curve placement. One benefit of this type of flow is the fact that no global constraints are placed on the image. Thus, the foreground and background can be heterogeneous and a correct segmentation can still be achieved in certain cases.

The Region-based approaches model the foreground and background regions statistically and find an energy optimum where the model best fits the image. Some region-based active contour models assume the various image regions to be of constant intensity. There are many advantages of region-based approaches when compared to edge-based methods including robustness against initial curve placement and insensitivity to image noise. However, techniques that attempt to model regions using global statistics are usually not ideal for segmenting heterogeneous objects.

Within this framework, segmentations are not based on global region models. Instead, the foreground and background were to be described in terms of smaller local regions, removing the assumption that the foreground and background regions can be represented with global statistics.

6.4.1 Output of Active Contours Method

Figure 10: Segmentation using Active Contours Method

CHAPTER - 7

FEATURE EXTRACTION

7.1 INTRODUCTION

If the input data to an algorithm is too large to be processed and not much information in the data then the input data will be transformed into a reduced representation set of features (also named features vector). Transforming the input data into the set of features is called feature extraction. Feature extraction involves simplifying the amount of resources required to describe a large set of data accurately.

Feature extraction makes the analysis of the particular application easier. There are different types of algorithms based on the application. Features can be extracted based on the

Statistical parameters

Textural parameters

Statistical parameters are the features where the statistical features are considered such as the mean, variance, standard deviation, skewness and kurtosis. The textural parameters take the texture of the image into account and the features are extracted.

7.2 CALCULATION OF STATISTICAL PARAMETERS

1. Mean: This is defined as the average or mean of the matrix elements. It is given by the formula,

N

∑ Xi / n

i =1

2. Standard deviation: This is the standard deviation of the matrix values. It is given by,

σ =√µ (z)

3. Skewness: Skewness is a measure of the asymmetry of the data around the sample mean. If skewness is negative, the data are spread out more to the left of the mean than to the right. If skewness is positive, the data are spread out more to the right. The skewness of the normal distribution (or any perfectly symmetric distribution) is zero.

The skewness of a distribution is defined as,

y = E(x-µ)3

σ 3

4. Kurtosis: Kurtosis is a measure of how outlier-prone a distribution is. The kurtosis of the normal distribution is 3. Distributions that are more outlier-prone than the normal distribution have kurtosis greater than 3; distributions that are less outlier-prone have kurtosis less than 3.

The kurtosis of a distribution is defined as

y = E(x-µ)4

σ 4

7.2.1 Range for 1st order Statistical Features

KURTOSIS

SKEWNESS

MEAN

STANDARD DEVIATION

RMS VALUE

NORMAL

IMAGES

40.67-136.34

5.2-10.9

32.91-63.18

21.71-42.3

39.42-76.08

CYST IMAGES

6.83-31.18

1.534-4.809

57.42-121.86

21.97-55.08

52.56-129.45

MRD IMAGES

10.96-70.71

2.12-7.25

54.07-110.49

21.54-44.74

62.79-117.89

Table 2 :First order Statistical feature ranges

After calculating the statistical parameters, we see that there is a distinct variation between the normal and Abnormal cases.

7.3 TEXTURAL PARAMETERS-GLCM ALGORITHM

GLCM stands for Gray Level CoOccurence Matrix.

GrayCoMatrix creates a gray-level co-occurrence matrix (GLCM) from image. GrayCoMatrix creates the GLCM by calculating how often a pixel with gray-level (grayscale intensity) value i occurs horizontally adjacent to a pixel with the value j. Each element (i,j) in GLCM specifies the number of times that the pixel with value i occurred horizontally adjacent to a pixel with value j.

By the GLCM toolbox 22 features can be obtained, which are listed below:

Autoc : Autocorrelation

Contr : Contrast

Corrm : Correlation Matlab

Corrp : Correlation [1, 2]

Cprom : Cluster Prominence

Cshad : Cluster Shade

Dissi : Dissimilarity

Energ : Energy

Entro : Entropy

Homom : Homogeneity Matlab

Homop : Homogeneity [2]

Maxpr : Maximum Probability

Sosvh : Sum of Squares Variance

Savgh : Sum Average

Svarh : Sum Variance

Senth : Sum Entropy

Dvarh : Difference Variance

Denth : Difference Entropy

Inf1h : Information Measure of Correlation 1

Inf2h : Information Measure of Correlation 2

Indnc : Inverse Difference Normalized (INN)

Idmnc : Inverse Difference Moment

Normalized

Co-occurrence matrix were calculated in 4 angles (0, 45, 90 and 135) and 2 distances (1and 3) and mentioned features were obtained in each angle/distance pair. Those features that showed great variation were used for further analysis.

7.3.1 Range for GLCM Textural Features

AUTOC

CONTR

CPROM

CORRM

NORMAL

IMAGES

3.95-8.50

0.28-0.47

10.62-102.03

0.70-0.88

CYST IMAGES

5.76-33.74

0.15-0.39

13.46-95.2

0.78-0.95

MRD IMAGES

6.35-33.7

0.13-0.41

11.61-89.97

0.77-0.93

Table 3 :GLCM based Textural feature ranges-1

HOMOM

DVARH

DENTH

ENERG

ENTRO

NORMAL IMAGES

0.89-0.93

0.21-0.37

0.40-0.57

0.18-0.43

1.40-2.14

CYST IMAGES

0.90-0.95

0.11-0.33

0.32-0.51

0.15-0.42

1.28-2.33

MRD

IMAGES

0.88-0.96

0.07-0.34

0.26-0.57

0.14-0.45

1.33-2.37

Table 4 :GLCM based Textural feature ranges-2

7.4 ANOVA TEST

We performed ANOVA test to determine the most significant feature that helps in correct classification of the ultrasound kidney images into the three categories when the features are fed into the classifier.

7.4.1 PTest for 1st order Statistical Features

FEATURE

NORMAL

CYST

MRD

KURTOSIS

6.17494E-08

3.82512E-09

6.58996E-07

SKEWNESS

0.007827

1.6E-06

5.96E-05

MEAN

1.22E-06

7.37E-10

9.06E-09

DEVIATION

8.78E-10

9.45E-11

4.17E-10

RMS

0.015997027

4.23081E-10

9.44491E-06

Table 5 :ANOVA Test Results for 1st order Statistical Features

PTest for Textural Features

FEATURE

NORMAL

CYST

MRD

AUTOC

4.55E-10

2.97E-08

1.06E-09

CONTR

0.000166

0.00024

1.62E-05

CPROM

0.000182

0.000274

1.94E-05

CORRM

3.74E-06

0.000259

7.79E-05

HOMOM

8.37E-16

2.39E-23

9.6E-13

DVARH

3.74E-13

2.43E-16

2.07E-13

DENTH

6.73E-06

3.25E-06

0.000494

ENERGY

2.44E-11

7.59E-15

2.53E-10

ENTRO

5.45E-11

1.14E-11

1.85E-11

Table 6 :ANOVA Test Results for Textural Features

According to the ANOVA test, the feature, DEVIATION is the most significant amongst 1st order statistical features & the feature, HOMOGENEITY is the most significant amongst textural features for discrimation between the three classes.

CHAPTER -8

CLASSIFICATION TECHNIQUES

8.1 INTRODUCTION

After processing the images, all the values and results need to be classified. When there is large amount of data present, checking if the given image is normal or abnormal individually is a tedious task. In order to make this easier classification algorithms are used.

For this particular application, Statistical (Bayesian) and Artificial Neural Network Algorithms are used. Three types of Neural Network algorithms are used, namely, MLP (Multi Layer Perceptron), SVM (Support Vector Machine), PNN (Probabilistic Neural Network).

8.2 BAYESIAN CLASSIFIER

A Bayes classifier is a simple probabilistic classifier. It is based on applying Bayes' theorem with strong independence assumptions. The underlying probability model would also be referred as "independent feature model". A naive Bayes classifier is based on the assumption that the presence of a particular feature of a class is not related to the presence of any other feature. On account of the precise nature of the probability model, Bayes classifiers are trained very efficiently in a supervised learning setting. Despite their naive design, over-simplified assumptions, Bayes classifiers have worked really well in many complex real-world situations. An important advantage of the naive Bayes classifier over others is that it requires little amount of training data to estimate the parameters necessary for classification.

The Bayes Classifier classifies data in two steps:

Training step: Using the training samples, the method estimates the parameters of a probability distribution, assuming features are conditionally independent given the class.

Prediction step: For any unseen test sample, the method computes the posterior probability of that sample belonging to each class. The method then classifies the test sample according the largest posterior probability.

The class-conditional independence assumption simplifies the training step because the one-dimensional class-conditional density for each feature can been estimated individually. While the class-conditional independence between features is not valid in most of the cases, it has been shown that this assumption yields good results in practice. This assumption allows the classifier to better estimate the parameters required for classification and uses less training data than other classifiers. This classifier is based on estimating P(X|Y), the probability or probability density of features X given class Y. Bayes Classifier provides support for normal (Gaussian), kernel, multinomial, and multivariate multinomial distributions.

8.2.1 Outputs for Bayesian Classifier

Figure 11: Distribution of feature values in the data set images

Figure 12: Misclassified samples-LDA

Figure 13: Misclassified samples-QDA

Figure 14: Misclassified samples-Gaussian Distribution

Figure 15: Misclassified samples-Kernel density

8.2.2 Results for Bayesian Classifier

Figure 16: Results for Bayesian Classifier

8.3 MULTILAYER PERCEPTRON

8.3.1 Introduction

A Multilayer perceptron is a network architecture which has been formulated with two adaptive parameters, the scaling and translation of the post synaptic function at each node. Artificial neural networks contain great number of processing elements which are connected to each other; the strengths of the connections are called weights. An MLP neural network is generally used for modeling the physical systems. It consists of a layer of input neurons, a layer of output neurons and one or more hidden layers. The optimal method of finding the number of hidden layers is by trail and error. For this application, two layer networks are used with 1 hidden layer (with 20 nodes). In ANN's, the knowledge lies in the interconnecting weights between neurons. The training process is an important characteristic of the ANN methodology. There are number of training algorithms used to train a MLP and the most frequency used one is the back-propagation training algorithm.

Figure17: MLP Network

Type of Neural Network

Number of Hidden Neurons

Activation Function

Learning Algorithm

Number of Epochs

Multilayer Perceptron

20

TansigMoid, Purelin

Levenberg Marquardt

10

Table 7: Neural Network Configuration

8.3.2 Training Procedure

The first step in developing a neural network is to create a database of examples. The examples will contain all the types of data which has to be classified. For this application we are using a multi feature classification. Here, many features are considered to classify the given data into the right category. The neural network used here is fully interconnected architecture. All this data is now fed to the system. This procedure is called training the neural network. By training the system, the system will familiar with the type of data given during the testing process and categorize them properly. The number of data used for this application is 50 (normal, cyst and medical renal disease data). Features used are mean, standard variation, skewness, kurtosis, autocorrelation, contrast, correlation, cluster Prominence, homogeneity, energy, entropy, difference energy, difference entropy, and wavelet decomposition- horizontal and vertical coefficients.

8.3.3 Testing Procedure

For testing purpose, a set of data (20) is set aside. After the training procedure, the data is given to the artificial neural network system. Based on the training efficiency, the data given for testing will be classified. After all the classification is done the efficiency of that system is known. For this application, Artificial Neural Network Tool box is used. This makes the implementation of the process easier.

Figure 18: Neural network main window

8.3.4 Curves & Plots for MLP Classifier

Figure 19: Validation curve

Figure 20: Plot for training

8.3.5 Results for MLP Classifier

Figure 21: Results for MLP Classifier

For running the Multilayer perceptron Artificial Neural Network we found that there is 90 % efficiency (multiple features) in the system to classify Ultrasound Kidney Images into Normal, Cyst and MRD (Medical Renal Disease) categories.

8.4 SUPPORT VECTOR MACHINE

8.4.1 Introduction

Support vector machines (SVMs) are a set of related supervised learning methods used for classification and regression. In simple words, given a set of training examples, each marked as belonging to one of two categories, an SVM training algorithm builds a model that predicts whether a new example falls into one category or the other. Intuitively, an SVM model is a representation of the examples as points in space, mapped so that the examples of the separate categories are divided by a clear gap that is as wide as possible.

SVM starts separating the data with a hyper plane and then extends this procedure to non-linear decision boundaries using the kernel trick. A classification task usually involves separating data into training and testing sets. Each instance in the training set contains one target value" (i.e. the class labels) and several attributes" (i.e. the features or observed variables). The goal of SVM is to produce a model (based on the training data) which predicts the target values of the test data given only the test data attributes.

8.4.2 Training Procedure

The first step in developing a neural network is to create a database of examples. The examples will contain all the types of data which has to be classified. For this application we are using a multi feature classification. Here, many features are considered to classify the given data into the right category. The neural network used here is fully interconnected architecture. All this data is now fed to the system. This procedure is called training the neural network. By training the system, the system will familiar with the type of data given during the testing process and categorize them properly. The number of data used for this application is 57 (normal, cyst and medical renal disease data).

8.4.3 Testing Procedure

For testing purpose, a set of data is set aside. After the training procedure, the data is given to the artificial neural network system. Based on the training efficiency, the data given for testing will be classified. After all the classification is done the efficiency of that system is known. For this application, Artificial Neural Network Tool box is used. This makes the implementation of the process easier.

8.4.4 Outputs for SVM Classifier

Figure 22: Kurtosis & Contrast (SVM)

Figure 23: Kurtosis and Dentropy (SVM)

Figure 24: Energy and Mean (SVM)

8.4.5 Results for SVM Classifier

FEATURES

PERCENTAGE CORRECT CLASSIFICATION

Kurtosis & Contrast

100.00

Kurtosis & Dentropy

85.71

Skewness & Homogeneity

100.00

Dvariance & Skewness

85.71

Energy & Mean

71.43

Contrast & Cprom

85.71

Table 8: SVM Classification

PROBABILISTIC NEURAL NETWORK

8.5.1 Introduction

Probabilistic networks do classification, in which the target variable is categorical. It is a kind of radial basis network that can be used for classification problems.

The comparison of PNN with MLP:

It is much faster to train a PNN network than a MLP network.

PNN networks are more accurate than MLP networks.

PNN networks produce predicted target probability scores.

PNN networks achieve Bayes optimal classification.

PNN network requires extra memory space to store the model.

In this algorithm a two-layer network is created. The first layer has radbas neurons, and the weighted inputs are calculated with dist and net input with netprod. The second layer consists of compet neurons, and its weighted input is calculated with dotprod and the net inputs with netsum. Only the first layer has biases.

8.5.2 Training Procedure

The first step in developing a neural network is to create a database of examples. The examples will contain all the types of data which has to be classified. For this application we are using a multi feature classification. Here, many features are considered to classify the given data into the right category. The neural network used here is fully interconnected architecture. All this data is now fed to the system. This procedure is called training the neural network. By training the system, the system will get familiar with the type of data given during the testing process and categorize them properly. The number of data used for this application is 57 (normal, cyst and medical renal disease data all together).

8.5.3 Testing Procedure

For testing purpose, a set of data is set aside. After the training procedure, the data is given to the artificial neural network system. Based on the training efficiency, the data given for testing will be classified. After all the classification is done the efficiency of that system is known. For this application, Artificial Neural Network Tool box is used. This makes the implementation of the process easier.

8.5.4 Outputs for PNN Classifier

Figure 25: Three vectors - Training (PNN)

Figure 26: Classification of new test set

8.5.5 Results for PNN Classifier

Figure 27: Results for PNN Classifier

CHAPTER - 9

CONCLUSION

In the pre-processing stage, amongst the various filters used, Median filter was found to be best suited for removing speckle noise. Once the noise was removed and the kidney region segmented, 1st order statistical and textural features were extracted. ANOVA test was performed on the entire set of features and the features DEVIATION and HOMOGENEITY were found to be the most significant features that aid in correct classification.

The classification efficiency of Naïve Bayes' classifier was highest for features Kurtosis & Contrast which is 89.80% when Gaussian distribution is used and 91.84% when Kernel distribution is used.

The classification efficiency of MLP was found to be 90% for multi features in discriminating the three classes of the test images.

The SVM classifier yielded the highest efficiency of 100% correct classification, for the set of features, Kurtosis & Contrast.

The Probabilistic Neural Network yielded 85% efficiency for multi feature input in differentiating between the three classes of test images.

From the above analysis, we conclude that from 1st Order Statistical parameters and Textural features, we can identify whether the given Ultrasound Kidney image is Normal or Kidney with Cyst or Kidney with Medical Renal Disease (MRD) using various classifiers appropriately.

APPENDIX

OUTPUTS OF VARIOUS FILTERS

Arithmetic Mean Filtered Image

Midpoint Filtered Image

Max Filtered Image

FEATURE EXTRACTION

1st Order Statistical Features Tabulation-Normal images

Textural Features Tabulation-Normal Images

1st Order Statistical Features Tabulation-Cyst images

Textural Features Tabulation -Cyst Images

1st Order Statistical Features Tabulation-MRD images

Textural Features Tabulation -MRD Images

Writing Services

Essay Writing
Service

Find out how the very best essay writing service can help you accomplish more and achieve higher marks today.

Assignment Writing Service

From complicated assignments to tricky tasks, our experts can tackle virtually any question thrown at them.

Dissertation Writing Service

A dissertation (also known as a thesis or research project) is probably the most important piece of work for any student! From full dissertations to individual chapters, we’re on hand to support you.

Coursework Writing Service

Our expert qualified writers can help you get your coursework right first time, every time.

Dissertation Proposal Service

The first step to completing a dissertation is to create a proposal that talks about what you wish to do. Our experts can design suitable methodologies - perfect to help you get started with a dissertation.

Report Writing
Service

Reports for any audience. Perfectly structured, professionally written, and tailored to suit your exact requirements.

Essay Skeleton Answer Service

If you’re just looking for some help to get started on an essay, our outline service provides you with a perfect essay plan.

Marking & Proofreading Service

Not sure if your work is hitting the mark? Struggling to get feedback from your lecturer? Our premium marking service was created just for you - get the feedback you deserve now.

Exam Revision
Service

Exams can be one of the most stressful experiences you’ll ever have! Revision is key, and we’re here to help. With custom created revision notes and exam answers, you’ll never feel underprepared again.