Please use this identifier to cite or link to this item: http://hdl.handle.net/10266/6025
Title: Computer Aided Diagnosis of Liver Diseases Using Ultrasound Images
Authors: Bharti, Puja
Supervisor: Mittal, Deepti
Keywords: Ultrasound;Feature extraction;Image Processing;Liver diseases;Computer-aided diagnosis;image enhancement;classification
Issue Date: 23-Sep-2020
Abstract: Liver is one of the most important organs in the human body, because it plays a vital role in digestion and in the metabolization of protein, drugs, toxins etc. Diseases of the liver are becoming major cause of morbidity and mortality all over the world. Ultrasound is preferable imaging modality by medical practitioners for the screening of liver diseases. It is preferable because of its several advantages; it is a non-invasive, cost effective, non-ionizing, portable and real time imaging modality. Medical practitioners/radiologists diagnose liver diseases by the visual examination of ultrasound images. Visual examination is a subjective criterion and is highly dependent on the expertise of radiologists in the domain. This may lead to ambiguity in the diagnostic procedure. To improve the objectivity in the diagnostics of liver diseases, various approaches of designing computer-aided support system are explored in this thesis. In addition, the visualization of liver diseases in ultrasound images becomes a tedious task due to overlapping textural characteristics of liver diseases. Thus, to design the computer-aided diagnostic (CAD) system, in-depth textural analysis of ultrasound images is performed to quantify information related to liver diseases. This research work is carried out with a database of 189 B-mode ultrasound images. The database was developed by collecting images from patients visiting Manipal Hospital, Bangalore, India during the period from March 2013 to August 2014. These patients were recommended for liver examination. In this duration of one year, ultrasound images were collected from 94 patients; among those 67 were male (age range: 21-70) and 27 were female (age range: 23-61). The database encompasses 48 images of normal liver and 141 images of abnormal liver. Images related to abnormal liver comprise of 50 images of chronic liver, 50 images of cirrhotic liver, and 41 images related to HCC evolved over cirrhosis. Diagnostically relevant areas in liver images are regions-of-interest (ROIs). The ROIs were identified and marked by the radiologist in all the images. Regions-of-interest were segmented into maximum possible square shape regions to create a large dataset for research work. These segmented regions are termed as segmented-regions-of-interest (SROIs). In this research work, three datasets are formed to optimize the design of CAD system. First dataset consists of 754 SROIs of size 32 x 32 pixels and is termed as original1 dataset. Second dataset consists of 400 SROIs of size 227 x 227 pixels and is termed as original2 dataset. Third dataset consists of ultrasound images on which enhancement was performed. This dataset is termed as enhanced dataset and has 754 SROIs of size 32 x 32 pixels. Ultrasound images suffer from low contrast and inherent formation of speckle. This may affect visual evaluation of liver diseases by radiologist. Therefore, in this research work, an image enhancement method is designed to reduce speckle and improve contrast of ultrasound images without the loss of diagnostic information. The proposed method is based on the scaling with neutrosophic similarity score (NSS), where an image is represented in the neutrosophic domain through three membership subsets; degree of truth (T), indeterminacy (I) and falseness (F). The NSS measures the belonging degree of pixel to the texture and is calculated with multicriteria: intensity, local mean intensity and edge detection. Then, NSS is utilized to extract the enhanced coefficient and this coefficient is applied to scale the input image. This scaling reflects contrast improvement and denoising effect on ultrasound images. The performance of proposed enhancement method is evaluated using both subjective and objective criteria. Results demonstrate that the proposed enhancement method enhances the contrast of ultrasound images and thus making the visualization of diagnostic information easier for the radiologist in comparison to original images. In ultrasound images, echo pattern of liver represents the texture of liver tissues. Hidden information in texture can be extracted using various mathematical descriptors to quantify information. These mathematical descriptors are termed as features and play a crucial role in differentiating liver tissue types. The choice of features also has an important influence on theaccuracy of learning algorithm and the time required for execution. In this research work, features are obtained from traditional hand-designed methods and deep learned methods. The features obtained via traditional hand-designed methods are termed as handcrafted features. On the other hand, the features obtained via deep learned methods are termed as deep learned features. On the basis of extensive literature review, the traditional hand-designed methods chosen for this research work are (i) gray-level difference matrix (GLDM), (ii) gray-level co-occurrence matrix (GLCM) and (iii) ranklet transform (ranklets). GLDM method relies on the calculation of local derivatives between pair of gray-levels. These local derivatives can provide information about the roughness of liver surface. In literature, the derivative calculation of GLDM method is based on the absolute differences between two neighboring pixels. However, the use of pixel pair, only in calculation of local derivatives, may not be able to provide the best approximation of calculated features from GLDM. Hence in this research work, GLDM features are calculated with better approximation of derivatives by using greater than two pixels. This proposed extension of GLDM is designed by computing the absolute difference of intensities using three, five and seven pixels. A total set of 148 features (64 of GLDM, 48 of GLCM and 36 ranklets) were extracted from handcrafted methods. However, this feature space may contain redundant features and the relevance of features in discriminating among liver tissue types is not known in advance. Thus, a hybrid feature selection method is designed in this research work. This method selects the optimal set of features to reduce computational complexity and improve performance. Further, the features extracted from methods such as GLDM and its extension, GLCM and ranklets may contain complementary information. To take the benefit of this information, feature fusion schemes (i) serial feature combination, (ii) serial feature fusion and, (iii) hierarchical feature fusion are also implemented in this research. The deep learned features extracted from deep neural network can be beneficial as they present the possibility of learning complex patterns from a large dataset. The hidden layers inside the deep neural network architecture automatically extract features without user intervention to give the best texture discriminating performances in a multi-class texture classification problem. In computer vision, convolutional neural networks (CNNs) are one of the most successful feature learning methods in performing various tasks. In this research work, two major approaches that employ CNN for feature extraction to classify liver diseases: (i) pre-trained CNN, and (ii) finetuning with transfer learning, are utilized. In the first approach, the convolutional base is kept in its original form and then its outputs are used to feed the classifier. Here the pre-trained AlexNet model is used and features are extracted from its different layers. In the second approach, the pretrained AlexNet model is adapted to the current research problem by replacing the last fully connected layer (intended for 1000 classes) with a new fully connected layer for four classes. The initial filter weights of CNN are derived from ImageNet dataset and then fine-tuned (optimized) through back-propagation so that they better reflect the four liver classes. The success of a computer-based system depends both on the features and classification method. An efficient set of textural features decides the correct detection of liver diseases and an appropriate classification method provides potential to produce accurate classification using these features. Thus, in this research an ensemble model is designed with different base classifiers: kNN, SVM, and RF; and combiner: majority voting. Independent training of all the base classifiers is done and finally, output of classifiers is combined with majority voting. The proposed ensemble model can be regarded as a combination of different strategies that enhances the generalization ability of model, as each component in the model learn about some part of classification problem. The CAD system designed by using this proposed ensemble model and handcrafted features extracted from original1 dataset gave 96.6% classification accuracy. Sensitivity and specificity of normal/chronic/cirrhosis/HCC were 96.3%/95.5%/97.5%/96.9% and 99.2%/98.0%/98.2%/99.8% respectively. The CAD system designed with proposed ensemble model and deep learned features extracted from original2 dataset gave 68.4% and 75.4% classification accuracy with pre-trained CNN and fine-tuned CNN. Here, it was observed that CAD system designed with handcrafted features gave better performance for classifying liver tissue type as compared to deep learned features. Finally, a CAD system was designed with handcrafted features and proposed ensemble model with enhanced dataset gave 97.9% accuracy which show the high precision capability of the system in predicting the test data to its actual liver tissue class. The experiment results also indicate the high performance of this proposed CAD system with the overall average sensitivity of 97.4% revealing its ability to perform well in predicting the true positive cases in each liver tissue type correctly. Sensitivity and specificity of normal/chronic/cirrhosis/HCC were 97.9/96.5/98/97.5% and 99.2/99.2/98.5/99.8% respectively. It was observed that there was 1.3% increase in accuracy with the CAD system designed using enhanced dataset in comparison to the CAD system designed with original1 dataset. Finally, it can be concluded that the CAD system designed to classify four liver tissue types comprises of (i) enhanced ultrasound images, (ii) hierarchal fused features set of handcrafted features and, (iii) ensemble classifier. The substantial performance of experiments on clinical database supports the strong candidature of the proposed computer-aided system to assist radiologist/clinicians in the diagnosis of liver diseases.
URI: http://hdl.handle.net/10266/6025
Appears in Collections:Doctoral Theses@EIED

Files in This Item:
File Description SizeFormat 
THESIS Puja Bharti.pdf2.78 MBAdobe PDFThumbnail
View/Open


Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.