Combined Statistical and Multiscale View on Ultrasonic Liver Images for Characterization

Abstract

The use of ultrasonography as an imaging modality has become widespread because of its ability to visualize the organs with no deleterious effects and low cost. Using ultrasonic liver images focal diseases can be identified by differences in echogenicity between normal and areas affected by diseases. In the presence of diffused disease, however, the entire organ may be affected. In that situation, there is no contrast in echo intensity on which to base a diagnosis. Hence it is difficult for an experienced clinician to diagnose the diffused liver diseases by simple visual interpretations. This can be improved by providing useful information, obtained by computer aided tissue characterization, that cannot be obtained by simple visual interpretation.Many researchers have studied the problem of liver tissue characterization. It is difficult to classify human body organ tissues using shape or gray level information because the shape of each organ is not consistent throughout all slices of a medical image and the gray level intensities overlap considerably for soft tissues. However, tissues are expected to have consistent and homogeneous structures, described as texture within liver images, making use of image texture analysis suitable for computer assisted characterization.Nicholas et al. (1) used textural features of the B-scan images to discriminate between the liver and spleen of normal humans. Paik and Fox (2) have proposed Hartley transform based texture measures to detect abnormal liver patterns. Chen et al. (3) found that the fractal dimension could be obtained in medical images by the concept of fractional Brownian motion (FBM) and were used to classify normal and abnormal ultrasonic liver images. Wu et al. (4) have proposed a multiresolution fractal (MF) based on the concept of multiple resolution imagery and FBM for distinguishing between hepatoma, cirrhosis, and normal liver, which produced an accuracy of 90%. The spatial gray level dependence matrix (SGLDM) features were proposed by Haralick et al. (5). Gebbinck et al. (6) applied discriminate analysis with supervised and unsupervised neural networks and tested their ability to detect various types of diffused liver diseases. This method has the ability to discriminate diseased from normal, but gives a decreased accuracy in simultaneous classification of all diseases. Kadah et al. (7) extracted the first-order gray level parameters such as mean and first percentile of the gray level distribution and second-order gray level parameters such as contrast, angular second moment, entropy, and correlation from the liver images and trained the functional neural network to classify normal, fatty, and cirrhosis liver. They showed that very good diagnostic rates can be obtained using unconventional classifiers trained on actual patient data.Mojsilovic et al. (8) have shown that the scale/frequency based on the separable wavelet transform is an appropriate feature extraction method for the analysis of ultrasound liver images. Mojsilovic et al. (9) investigated the application and advantages of the nonseparable wavelet transform features for liver tissue characterization and compared the approach with other texture measures like SGLDM, fractal texture measures, and Fourier measures. The classification accuracy was 87% for the SGLDM, 82% for Fourier measures, 69% for fractal, and 90% for the wavelet approach.Pavlopoulos et al. (10) extracted features such as fractal dimension texture analysis (FDTA), SGLDM, gray level difference statistics (GLDS), gray level run length (RUNLs) statistics and first-order gray level parameters (FOPs) from the ultrasonic liver images and the reduced optimal feature set was used to train the fuzzy neural network. The system produces an accuracy on the order of 82.6%. Lee et al. (11) proposed a feature selection algorithm based on fractal geometry and M-band wavelet transform for the classification of normal, cirrhosis, and hepatoma ultrasonic liver images. A hierarchial classifier, which is based on the proposed feature extraction algorithm, is at least 96.7% accurate in distinguishing between normal and abnormal liver images and is at least 93.6% accurate in distinguishing between cirrhosis and hepatoma liver images.Neural network classifiers have demonstrated promise in pattern classification and been considered as a potential alternative approach to statistical pattern classification of liver lesions. However, determining the number of hidden layers and number of hidden nodes in each layer is still a fundamental and open question that is often raised in applying multilayer neural networks to actual problems. In reality, they are typically determined experimentally (12). Furthermore, even a training set with zero error does not guarantee that the network can work well in actual cases (13).In this work, the SGLDM features are extracted from the detail images obtained by applying biorthogonal wavelet transform on the region of interest (ROI), selected from the ultrasonic liver images. The optimal feature set is used to train the probabilistic neural network (PNN) for the classification of normal, fatty, and cirrhosis. The system produces accuracy on the order of 92.4%.The aim of this work is to present an algorithm for the computer aided diagnosis (CAD) system to classify the liver as normal, fatty, and cirrhosis using image processing and neural network techniques. The proposed system structure is shown in Fig.1.The ultrasound images used in this study were captured by the SIEMEN ACUSON ASPEN ultrasound system using convex electronic array multifrequency high-density transducers operating from 1.5MHz to 4.0MHz. However, the images were acquired using transducers operating at 3.0MHz and 4.0MHz. The images are tissue harmonic images. The images were captured from 75 normal (healthy) persons, 75 fatty liver, and 75 cirrhosis liver patients, with 256 gray level resolutions.The square shaped ROI is selected from the liver images. The area is chosen with help of a radiologist so that the area contains only liver parenchyma.Feature extraction is the operation to extract various image features for identifying or interpreting meaningful physical objects from images. A weakness shared by statistical features is that the image is analyzed at one single scale. This limitation can be lifted by employing multiscale representations (1415). Texture classification experiments were performed by Mojsilovic et al. (16) using Haar filters, Daubechies filters, and eight different biorthogonal filter pairs and they reported that biorthogonal filters are more suitable for texture analysis. Hence, in this work, biorthogonal wavelet based statistical texture features are extracted from the ROI of the ultrasonic liver images. The selection of biorthogonal transform was made because of smoothness and their robustness under slight shifts of image components.The two-dimensional (2D) discrete wavelet transform is computed by applying a separable filter bank to the image as given by Eqs. 1,2,3,41Ln(bi,bj)=[Hx*[Hy*Ln−1]∣2,1]∣1,2(bi,bj)2Dn1(bi,bj)=[Hx*[Gy*Ln−1]∣2,1]∣1,2(bi,bj)3Dn2(bi,bj)=[Gx*[Hy*Ln−1]∣2,1]∣1,2(bi,bj)4Dn3(bi,bj)=[Gx*[Gy*Ln−1]∣2,1]∣1,2(bi,bj)where * denotes the convolution operator, (∣1,2)∣2,1 the subsampling along the rows (columns), and L0=I is the original image. H and G are a low and bandpass filter, respectively. Ln is obtained by lowpass filtering and is therefore referred to as the low-resolution image at scale n. The detail images Dni are obtained by bandpass filtering in a specific direction and contain directional detail information at scale n. The original image I is thus represented by a set of subimages at several scales: {Ld,Dni}i=1,2,3,n=1…d which is a multiscale representation of depth d of the image I. Since each wavelet coefficient Dni(bi,bj)∊R and the cooccurrence matrix is defined for an image with a countable number of gray levels, the cooccurrence matrix Pnidθ can be defined. The element (j,k) of the cooccurrence matrix Pnidθ is defined as the joint probability that a wavelet coefficient Dni=j cooccurs with a coefficient Dni=k on a distance d in direction θ. In this work 1 is used for d and 0deg, 45deg, 90deg, and 135deg are used for θ. Pn1dθ represents horizontal details of the image; Pn2dθ represents vertical details of the image; and Pn3dθ represents diagonal details of the image where n represents the scale of the wavelet transform.The 14 measures of textural features (5), angular second moment (ASM) or energy, contrast, correlation, sum of squares, or variance, inverse difference moment (IDM), homogeneity, sum average, sum variance, sum entropy, entropy, difference variance, difference entropy, information measures of correlation, and maximum correlation coefficient, are extracted from the cooccurrence matrix Pnidθ and averaged for d=1 and θ=0deg, 45deg, 90deg, and 135deg. Hence for each scale, from these detail images, 42 features are extracted.To reduce the cost of classifier, the extracted features are reduced by sequential backward selection while maintaining classification accuracy. Sequential backward selection begins with all features and repeatedly removes a feature whose removal yields the maximal performance improvement. The selected feature set is used as an input to the PNN for characterization.The PNN was developed by Specht. This particular type of artificial neural network (ANN) provides a general solution to pattern classification problems by following the probabilistic approach based on the Bayes formula (17). The key advantages of PNN are that training requires only a unique pass and that the decision hypersurfaces are guaranteed to approach the Bayes optimal decision boundaries as the number of training samples grows. The main criticism of PNN is that all training samples must be stored and used in classifying new patterns. To reduce the computational cost, dimensionality reduction and clustering methods are usually applied, previous to PNN construction. The architecture is given in Fig.2. It consists of three layers: a layer of input units, a layer of pattern units, and a layer of output units. The number of input units in the input layer is equal to the number of features. There is one to one correspondence between pattern units and training examples. So, pattern units are equal to training examples in numbers. Each output unit corresponds to a class. The PNN designed consists of 18 input units (I1–I18) in the input layer, 120 pattern units corresponding to 40 normal liver training examples (N1–N40), 40 fatty liver training examples (F1–F40), and 40 cirrhosis training examples (C1–C40). The number of training samples are chosen experimentally (incremented by ten for each testing) based on optimal classification performance. The output layer consists of three output units corresponding to normal liver, fatty liver, and cirrhosis liver.The PNN is constructed as training progresses. The algorithm is given as follows.Step 1: Weight initializationThe input weight vector of pattern unit i is initialized to the ith pattern (training example) vector, which is normalized to unit length. The input weight vector of every output unit is initialized to a vector of all 1s.Step 2: Calculation of activation1The activations of input units are determined by the pattern (training example) presented to the network. The input vector is normalized to unit length.2The activation Oj of pattern unit j is given by5Oj=exp[(∑iWjiXi−1)σ2]where Wji is the connection weight from input unit i to pattern unit j and Xi is the activation of input unit i.3The activation Oj of output unit j is given by6Oj=1mP(wj)∑iWjiOiwhere Wij is the connection weight from pattern unit i to output unit j; m is the number of training examples labeled class wj which corresponds to output unit j; and P(wj) is the prior probability of wj. Oi is the activation of pattern unit i.4The decision of the network is the class corresponding to the output unit with maximum activation.The net can be used for classification as soon as an example of a pattern from each of the three classes has been presented to it. However, the ability of the net to generalize improves as it is trained on more examples. For classifying the new pattern, the features extracted are given to the input layer. The activation at the output unit is calculated. The class corresponding to the maximum activation is the class of the new pattern.PNN is not designed to be an incremental learner since old examples are represented explicitly in the network. Every new example demands the modification of the network architecture by the addition of another pattern unit and associated connections.All the design steps are implemented using MATLAB. From ultrasound liver images (Fig.3a), a square shaped region of interest (Fig. 3b) is selected with the help of radiologist using the function “imcrop” in MATLAB.Biorthogonal transform is applied on the ROI of the liver image to obtain horizontal, vertical, and diagonal detailed images. For these images, three SGLDMs are constructed. From each of these matrixes, 14 texture features and hence 42 total features are extracted. The feature set is reduced by sequential backward selection.From the experiments conducted, it is found that the optimal feature subset, which gives better classification performance, is the angular second moment, contrast, entropy, correlation, sum of squares (variance), and homogeneity.One might expect the performance to be improved, or at least unchanged, by making additional measurements on the images. But in this application domain, the overall detectability may decrease by including a correlated feature with marginal or no discriminability. In other words, since every element of a feature has its contribution to the classification performance, it will be degraded when the added feature elements have bad geometrical distribution, i.e., the distribution spreads over the feature space. This behavior is shown by the experiment conducted for feature selection.The sample optimal feature vector for normal, fatty, and cirrhosis liver extracted from horizontal detailed, vertical detailed, and diagonal detailed images are given in Tables 123, respectively. The extracted feature set is not redundant. Because the horizontal detailed image contains horizontal edge information of ROI, the vertical detailed image contains vertical edge information of ROI and the diagonal detailed image contains diagonal edge information of ROI.The 18 features extracted are formed as a feature vector for training the PNN for classification. The training set and testing set are disjoint. The training set contains 40 normal, 40 fatty, and 40 cirrhosis livers. By using this training set, PNN was constructed. The testing set which contains 35 samples in each class was used to test the PNN for identification of liver as normal, fatty, and cirrhosis. The performance of the proposed system was tabulated for the training set in Table 4.The reason for the misclassification may be due to the presence of blood vessels in the selected region of interest. The blood vessels may destroy the homogeneous structure of the image. The selection of ROI is constrained by the presence of blood vessels and acoustic shadowing. If ROI is selected, including blood vessels and acoustic shadowing, the liver sample is misclassified. The training set is correctly classified. Representative samples also have to be carefully chosen for good classification accuracy. The results obtained in this study and the previous result (711) clearly illustrate that the location of the ROI within an image has a dominant effect on the classification. For training, the data set is chosen as homogeneous as possible with the help of the radiologist.The same set of six features was extracted directly from the image without applying biorthogonal wavelet transform. The features are given as input to PNN for classification. The performance of the system for the same data set is given in Table 5.From Tables 45, it is clear that the wavelet domain improves the accuracy of liver tissue characterization. Classification accuracy of the feature set in the wavelet domain for scales=1, 2, 3, and 4 and gray level domain are analyzed for the choice of decomposition level and the performance measures are tabulated in Table 6.From Table 6, it is clear that the combined statistical and multiscale view on image improves the classification accuracy. The overall classification accuracy is calculated as given by Eq. 7 and tabulated in Table 7.7Classificationaccuracy=numberofcorrectlyclassifiedliverimages∕totalnumberofliverimagesFrom Table 7 it is clear that, when the scale for wavelet transform is increased, accuracy is also increased. But the computational cost (time) is also increasing. Hence the scale should be chosen based on accuracy and computational cost. Scale 3 gives optimal performance. Hence, features are extracted at Scale 3 for characterization of liver tissues. The overall classification accuracy of the system is 92.4%. Training took 5s. The liver sample was tested within 1s. If the training set size is increased, the performance may be increased. The classification accuracy of the feature set is compared with other authors sets and tabulated in Table 8.The proposed feature set gives good classification accuracy. If a new sono machine is used to acquire images, then images should be acquired with a new machine and the extracted features should be used for constructing the PNN. Then we can use it for CAD. The results were also evaluated from the ground truth collected from the radiologist and the patients doctor. The results are satisfactory to the doctors.Results show that an automatic computer based system for diagnosing liver diseases can be constructed by using a combined statistical and multiscale view on image and PNN. The optimal feature set derived from SGLDM after applying biorthogonal transform is the angular second moment, contrast, entropy, correlation, homogeneity and sum of squares (variance) and are able to characterize the liver tissues successfully with an error less than 10%. The results show that in the future hardware and software implementation of tissue analysis function can be provided with ultrasound scanner machines for radiologists to give a second opinion. The proposed prototype can also be extended for other types of liver images and liver tissues provided that the feature vectors are re-evaluated and the probabilistic neural network is well trained. Since the proposed design follows the component based model, each block can be replaced by the best one in the future.