CLASSIFICATION OF B-MODE ULTRASOUND KIDNEY AND LIVER IMAGES

Abstract

Imaging technology revolutionized the area of diagnosis by being complimentary to the clinical diagnosis because of its non-invasive nature. With rapid advancements in imaging modalities, the clarity and the volume of diagnostic information being obtained by radiologists have made the role of medical professionals easier and the patient comfortable. However, the subjective analysis of images takes much time of radiologists and is prone to human error depending upon their expertise and experience. To overcome these limitations, researchers have been involved in developing computer algorithms for extracting diagnostic information from the images based on clinical inputs about the diseases. The imaging modalities such as computed tomography (CT), magnetic resonance imaging (MRI), ultrasound (US) and others have their advantages and disadvantages in general. Each modality has different perspective of information to offer about a particular disease. B-mode ultrasound imaging is preferred as an initial examination for soft tissue structures like liver, kidney, prostate, uterus etc. Its cost effective and portable nature is more suitable for extensive usage in countries like India. However, the quality of images being obtained by ultrasound is relatively of poor quality, which hinders the interpretation of images. The images obtained from echo based ultrasound imaging, basically have an interference pattern termed as speckle noise. From a radiologist’s perspective, though the speckle noise is a hindrance in their subjective analysis of images, it does contain necessary diagnostic information. From a researcher’s perspective, one who develop algorithms for objective diagnosis, speckle noise is an interference which is making the task difficult. The computer-aided diagnosis (CAD) system involve majorly pre-processing followed by analysis and other applications. The present work is about the development of computer-aided classification (CAC) systems for the classification of B-mode ultrasound images. It also includes exploration of possibilities to enhance the efficiency of a CAC system. Considering the significance and frequency of cases of kidney and liver, radiologist suggested to develop CAC systems for both the databases. Thus, the present work primarily comprises of two parts: One is concerned with kidney images and the other involves liver images. In the first part, two CAC systems are proposed for kidney images: 1. a. Classification of normal, medical renal disease (MRD) and cyst classes of kidney using B-mode ultrasound images is carried out. It also includes the possible contribution by the features extracted from the de-speckled images for the classification task. x 1. b. The cyst class recognition is comparatively easier among the three classes. So, to improve the classification of normal and MRD classes, a binary CAC system is proposed. To optimize the combination of texture features and the de-speckling methods for the classification, exhaustive experiments have been also carried out. In second part, two CAC systems for the liver images are proposed: 2. a. First, classification of normal liver and grades of fatty liver (mild, moderate and severe) using B-mode ultrasound images are considered. It is also been highlighted that regions of interest (ROIs) from diaphragm have a contribution to make for the classification task. 2. b. There is an overlap of image characteristics among severe fatty liver and cirrhotic liver. So, a binary CAC system is proposed for the classification of severe fatty liver and cirrhotic liver using B-mode ultrasound images. The impact of a classifier associated with the wrapper method of feature selection on the CAC system has been also evaluated. A generalized block diagram which can depict the outline of each of the above stated four CAC systems of the present work is shown below. Figure 1: A brief work-flow diagram The database used in the present work consisted of 35 kidney images comprising of 11 normal, 16 cysts and 8 MRD images and 69 liver images which includes 12 normal, 14 mild fatty liver, 14 moderate fatty liver, 13 severe fatty liver and 16 cirrhotic liver images. The ROIs are required when sufficient number of images are not available or/and area of interest is confined to a small region. In the present work, sufficient numbers of non-overlapping ROIs of size 32 × 32 pixels are extracted from each image by the experienced radiologist. From each ROI, various texture features are extracted to represent the ROI image in quantitative form so that it can be used to perform objective classification of ROI images. In the present work, the features used often in classification of ultrasound medical images are considered, namely, first order statistics (FOS) based features, moment invariant (MI) features, run-length matrix (RLM) features, gray-level co-occurrence matrix (GLCM) features and Laws’ texture energy (Laws) features. A category of gradient (Grad) based features, which has not been used so frequently, is also considered to evaluate its potential in the present work. xi In order to reduce computations in further processes and to remove redundant features without compromising on the efficiency of the CAC system, a feature selection process is considered. Differential evolution feature selection (DEFS), which is a recent wrapper method has been used in the present work. An optimization process has been carried out in prior to find out the values to be assigned to the parameters of DEFS, viz. desired number of features (texture features of ultrasound images), population size (number of ROI patterns) and number of iterations (repetitions required for the algorithm to evolve till a saturation is reached). Support vector machine (SVM) classifier is often considered in medical applications for the better performance it brings. To perform the last stage of the CAC system i.e. classification, one-against-one multi-class SVM classifier has been considered in the present work. To evaluate the performances, present work is carried out basically in two stages i.e. without feature selection and with feature selection. Overall classification accuracy (OCA) is considered as a performance measure in the first stage. For feature selection stage, the DEFS process is repeated 30 times to obtain 30 subsets. The subset which produced highest OCA among the 30 subsets is considered as one of the measures to evaluate the performance after feature selection. The OCAs obtained from 30 subsets are used to calculate average accuracy (standard deviation) (AASD), and is considered as another measure to show the reliability of a particular feature set for the classification task. In the present work, the methodologies employed in different stages of CAC systems (ROI marking, extraction of features, feature selection and classification) are same for the two databases of kidney and liver. Thereby, we could draw conclusions on commonalities and distinguishing characteristics of the stages of CAC system for the two databases. Brief description of CAC systems: CAC System 1: The radiologist suggested that the distinguishing characteristics between normal and MRD is constrained in the region of parenchyma of kidney and cysts are local in nature. The significant point being considered is the selection of ROI for the classification of normal, MRD and cyst classes of kidney using B-mode ultrasound images. Generally, the de-speckling methods are used for providing better visualization of images for the radiologists, which in turn help them in making diagnosis easier. To overcome the limitation of subjective diagnosis, computer-aided systems are being proposed by the researchers to provide objective assistance in diagnosis. The changes in the textural information by the de-speckling methods are considered. Texture features are extracted from the de-speckled images and the feature sets are xii concatenated in different combinations to enhance the potentiality of the CAC system. The concatenated RLM feature sets extracted from the ROIs of images de-speckled by Lee’s sigma and enhanced Lee methods have resulted in an AASD of 86.3(1.6). CAC System 2: Among the normal, MRD and cyst classes of kidney, distinguishing normal and MRD is more challenging, and hence in the current objective only those two classes are considered. To evaluate the performance of different texture features extracted from the images de-speckled by various methods in the classification task, six categories of texture features and eight de-speckling methods are considered. RLM features from the images de-speckled by Frost method gave an AASD of 87.0(2.9). CAC System 3: A CAC system is proposed for the classification of normal liver and grades of fatty liver i.e. mild, moderate and severe using B-mode ultrasound images. Radiologists consider the visibility of diaphragm along with the variations in liver texture in their subjective diagnosis. For the CAC system also, ROIs are considered from both within liver and diaphragm areas to obtain higher accuracy of the system. To combine the information of these two regions, ratio features, inverse ratio features and additive features are computed. The Laws ratio features have performed better with an AASD of 84.9(3.2). CAC system 4: In the line of grades of fatty liver, the advanced stage of severe fatty liver is cirrhotic liver. A CAC system is proposed for the classification of severe fatty liver and cirrhotic liver, wherein ROIs from the diaphragm area are not considered. In the present work, DEFS, a wrapper method is being used for the feature selection. As an another objective, two classifiers i.e. Naïve Bayes (NB) and K-nearest neighbour (KNN) classifiers are used along with DEFS algorithm to obtain different subsets of features. The concatenated set of first-order statistics and Laws feature subsets obtained from KNN-DEFS produced better AASD of 99.5(0.8). For kidney images, the CAC system 1 can be employed if normal, MRD and cyst classes are considered for classification. If the result of this system is not cyst, the CAC system 2 can be used to enhance the classification accuracy of normal and MRD. Similarly, for liver images, the CAC system 3 can be used for the classification of normal and grades of fatty liver (mild, moderate and severe). If the output of this system is severe fatty liver, then CAC system 4 can be utilized for further clarification among severe fatty liver and cirrhosis.