Processing and Analysis of Ultrasound Images for Tissue Characterization

Abstract

Medical imaging is a non-invasive process of visualizing the inner body tissues, organs and bones for diagnosis purpose. It is the fast growing branch of biomedical engineering. These days many imaging modalities are available for different applications. Ultrasound imaging is the most popular because it is non-ionizing, low cost, portable and real time imaging. The main applications of ultrasound imaging are visualizing the fetus development, diagnosis of prostrate and abdominal organs like kidneys, gallbladder and liver. Liver is one of the most important organs in the human body, because it controls many important metabolisms. Fatty liver disease (steatosis) is highly prevailing disease among all the liver diseases in India. The visual examination of ultrasound images, for diagnosis of fatty liver is subjective and less accurate in marginal cases. Fatty liver is a condition that occurs when the fat content of the ‘hepatocytes’ increases, resulting in variation of the texture of liver surface. Therefore, quantitative texture analysis may give crucial information which is otherwise difficult to extract by visual interpretation of ultrasound images. In this research work, a Computer Aided Diagnostic (CAD) method is proposed for the liver tissue characterization using texture analysis. The ultrasound imaging has one limitation, and that is speckle, which degrades the visual quality of ultrasound images and masks some fine details of tissue under observation. Therefore it is important to suppress this speckle before any computer aided image processing or analysis, while keeping the diagnostic information intact. To accomplish this task, a modified fourth order partial differential equation (fpde) based filter is proposed, which is adaptive to local ‘coefficient of variance’ in the 3x3 spatial window. To further increase the efficacy of the filter, ‘edge-map’ technique is used, which enhances the edges and fine details in the ultrasound image. The performance of this proposed image enhancement method is evaluated on synthetic and ultrasound images through visual analysis by the expert radiologists as well as through quantitative metrics. It has been found that proposed ultrasound image enhancement method outperformed other existing methods. Liver is very large tissue and to extract the features for quantitative analysis, a Region of Interest (ROI) is selected. The ROI is a sub part of the image and is used to avoid unwanted blood vessels, bile duct, and cyst in the analysis, Moreover, the use of ROI will reduce the computational time. Thus, the selection of an appropriate ROI is very crucial in CAD systems. In this work, experiments were conducted to find the appropriate ‘shape, size and location’ of ROI. After the analytical study and discussions with radiologists, it has been observed that a square shaped ROI having size of 30x30 pixels size is optimal, and it should be taken along (or near) the centre line of the image. To increase the reliability of CAD, multiple ROIs are selected from a single image and subsequently, the average value of each feature from multiple ROIs is computed and used for further analysis. A lot of research has been done for liver characterization through ultrasound imaging. Although there are number of quantitative techniques for liver characterization such as radio frequency (RF) signal back-scattered and attenuation analysis, elastography etc., but texture analysis is the most reliable technique. Many texture models like Spatial Grey Level Co-occurrence Matrix (SGLCM), Grey Level Difference Statistics (GLDS), First Order Statistics (FOS), Law’s Texture Energy Measure (TEM), Fourier Power Spectrum (FPS), Statistical Feature Matrix (SFM), Fractal Feature (FF), Grey Level Run Length Matrices (GLRM) have been used for Ultrasound Tissue Characterization (UTC). These texture models are used with different classification methods and, each method has its own advantages and disadvantages. The more accurate methods are computationally expensive, while the easier to compute methods are not so accurate. Thus, keeping in view of the above discussion, there is a need to go for a computer aided, quantitative, improved method that give fairly high accuracy with lesser computational load. In this research work, an information fusion based method is proposed to help the radiologists in characterizing the fatty liver more accurately and fast. This novel method uses highly discriminating texture features in a linear combination to give a discriminative index (DI) to differentiate the fatty liver and normal liver. The highest discriminative feature has the highest weightage to find DI. The threshold value of DI is selected in such a way that it gives 100% sensitivity, which means, that none of the fatty liver is miss-classified as normal liver. The overall accuracy of the proposed method is 95% with 100% sensitivity, which is better than other existing methods. Another limitation of most of the existing quantitative methods is that they are able to characterize the liver into normal liver or fatty liver only. However, for a physician it is more important to know the level of fat accumulation in the liver (fatty liver grading), to suggest a better treatment. Fatty liver is further characterized as mild fatty, moderate fatty and severe fatty, depending upon the fat deposition in the hepatocytes. Moreover, highly discriminative texture features are very sensitive to the grey level in the image. While acquiring the ultrasound image, usually the radiologist varies the Time Gain Compensation (TGC) settings for better visual analysis. This variation in machine settings causes change in average gray level among different ultrasound images. Therefore the classification methods which are using ‘grey level’ based texture features, are prone to errors. To overcome this drawback, the grey level based features are used after taking the ratio of their values from liver and kidney. A limited number of researchers like Webb et al., Soder et al., Minhas et al. have used the ratio of grey-level based features for liver tissue characterization, and termed it as hepato-renal index or ratio. They have used ratio of only two of grey-level based features (mean grey level intensity and contrast). These researchers have not considered any frequency or fractal based feature. However in the present study, five highly discriminative grey level based features are used after taking their hepato-renal ratio. Moreover, frequency and wavelet based features are also combined and a feature vector is formed of highly discriminative features. Finally a multiclass Support Vector Machine (SVM) based classification method is presented to characterize the normal and fatty liver and then, the further grading of fatty liver into mild, moderate and severe fatty liver. In the proposed method polynomial kernel of SVM is selected in such a way that it gives 100% sensitivity, which means, that none of the fatty liver is miss-classified as normal liver. The overall classification accuracy of the proposed method is 98.3% with 100% sensitivity. The present work is likely to contribute significantly in the area of computer aided medical image analysis and diagnosis. An experimental study has also been done to find optimal size of ROI for liver tissue characterization. A novel classification method, based on information fusion is proposed in this thesis, which gives a discriminative index DI to classify the normal and fatty liver. However the concept is universal and applicable to any other binary classification problem. The SVM based fatty liver characterization method can be useful in timely diagnosis of mortal liver disease. Finally, some suggestions based on observations and experiments, are presented to carry out further research work in this area.