ANALYSIS AND CLASSIFICATION OF B-MODE LIVER ULTRASOUND IMAGES

Abstract

The present research work has been carried out with an aim to enhance the diagnostic potential of conventional B-Mode ultrasound (US) imaging modality for diagnosis of liver diseases. To achieve this objective, the design and implementation of an interactive system for diagnosis of liver diseases using B-Mode liver US images is proposed in the present study. The research objectives for the present work were formulated keeping in view the needs of the radiologists, based on the practical difficulties faced by them in routine clinical practice. The study was conducted by collecting a comprehensive image database of 124 B-Mode liver US images with representative cases from each image class, acquired from the patients who underwent US examination at the Department of Radiodiagnosis and Imaging, PGIMER, Chandigarh, India during a period from March 2010 to March 2012. The image database comprises of 21 Normal (NOR), 16 Cirrhotic, 12 Cyst, 15 Hemangioma (HEM), 28 Hepatocellular Carcinoma (HCC) and 32 Metastatic Carcinoma (MET) liver images. Further bifurcation of Cyst, HEM and MET images into typical and atypical cases, and HCC cases into small HCC (SHCC) and large HCC (LHCC) cases is shown in Fig. 1. Fig. 1 The description of image database used in the present research work. Note: FLLs: Focal liver lesions; HEM: Hemangioma; HCC: Hepatocellular carcinoma; MET: Metastatic carcinoma or metastasis; SHCC: Small HCC; LHCC: Large HCC. The proposed interactive system for diagnosis of liver diseases using B-Mode US images consists of two modules as shown in Fig. 2. Module 1 is designed to assist or provide second opinion to the radiologist if there is confusion within Normal, Cirrhosis, HCC or MET liver image classes. xxii Fig. 2 Block diagram of the proposed interactive system for diagnosis of liver diseases. Note: IROIs: Inside lesion regions of interest; SROIs: Surrounding lesion regions of interest. Early diagnosis of liver cirrhosis using texture descriptors computed from regions of interest (ROI) extracted from conventional B-Mode liver US images is clinically significant as most of the cirrhotic patients are asymptomatic, and the biochemical tests like elevated liver enzyme detect cirrhosis at an advanced stage. It is worth mentioning that the patients with liver cirrhosis are at high risk of developing hepatocellular carcinoma (HCC, a primary malignant focal liver lesion), and cirrhosis is also the leading cause of portal hypertension. xxiii Since fibrosis is an essential stage leading to liver cirrhosis, it is important to know whether the fibrotic changes in liver parenchyma correspond to liver cirrhosis so that medication can be administered timely and the associated complications can be prevented. Furthermore, since liver cirrhosis is considered as a pre-cursor to development of HCC and it is considerably difficult to diagnose small HCCs developed on already nodular cirrhotic liver parenchyma, the diagnosis as to whether the textural changes in the liver parenchyma are cirrhotic changes or they indicate the development of HCC is absolutely necessary. It is worth mentioning that there is a considerable overlap between the sonographic appearances of HCC and MET lesions, at the same time the differential diagnosis between HCC and MET lesions is essential for effective treatment of liver malignancies. Accordingly, the Module 1 of the proposed interactive system for diagnosis of liver diseases incorporates three different CAD systems: (i) CAD System-I for binary classification between normal and cirrhotic liver tissue, (ii) CAD System-II for classification between normal, cirrhotic and HCC liver, and (iii) CAD system-III for binary classification between HCC and MET liver malignancies. The region of interest (ROI) extracted by the radiologist, is fed to CAD system-I, for characterization between normal and cirrhotic liver tissue. Although the design of the proposed CAD system-I yields 100 % accuracy for characterization between normal and cirrhotic liver tissue, but due to severely limited sensitivity of US for detection of small HCCs developed on cirrhotic liver, it is quite possible that the region of interest (ROI) belonging to cirrhotic liver may actually represent a HCC. Therefore, the ROI which is predicted as cirrhotic by the CAD system-I is passed through CAD System-II for characterization between normal, cirrhotic and HCC liver tissue. If the prediction of the CAD System-II for an ROI is cirrhosis, it gives greater confidence to the radiologist that the liver tissue is cirrhotic. However, if for a particular ROI the decision of the CAD System-II is HCC, it is advised to investigate whether the ROI belongs to a HCC lesion or a MET lesion because of their significant overlapping sonographic appearances. Although, both HCC and MET lesions represent malignant liver lesions, differential diagnosis between HCC and MET lesion is absolutely necessary for better management of the disease and adequate scheduling of treatment options. Therefore, the ROI, if predicted as HCC by CAD System-II, is passed through CAD System-III for binary classification between HCC and MET liver tissue. If it is predicted as HCC, it gives greater confidence to the radiologist that the ROI represents HCC, or else if the ROI is predicted as xxiv MET, it is recommended to pass the ROI to the CAD System-IV of Module 2 of the proposed interactive system for diagnosis of liver diseases for its differential diagnosis with other FLLs. Exhaustive experimentation was carried out for the design of efficient classifiers for CAD System-I, CAD System-II, CAD System-III and CAD System-IV. The radiologists observe the texture patterns of the IROIs (inside lesion regions of interest, i.e., regions of interest extracted from region well inside the lesion boundary) as well as the texture patterns of the SROIs (surrounding lesion regions of interest, i.e., regions of interest extracted from liver parenchyma surrounding the lesion and approximately at the same depth as that of the centre of the lesion) for making differential diagnosis between HCC and MET lesions using B-Mode US images. Accordingly, texture feature extraction from IROIs and SROIs were carried out for design of CAD System-III for classification between HCC and MET lesion. Therefore, if the prediction of CAD system-II for a ROI is HCC, the radiologist is required to mark an additional SROI for the classification task. The results obtained from exhaustive experiments carried out in the present work indicate that the texture feature extraction from both IROIs and SROIs enhances the efficiency of the CAD System-III for classification between HCC and MET lesions. The classification performance obtained by CAD System-I, CAD-System-II and CAD System-III designs implemented for the design of Module 1 of the proposed interactive system for diagnosis of liver diseases using B-Mode US images is depicted in Table 1. Table 1 Classification performance CAD System-I, CAD System-II and CAD System III of Module 1 CAD System Design OCA (%) CAD System-I: (Normal and cirrhotic Liver): Design based on first four singular value mean features derived by singular value decomposition of gray level co-occurrence matrix and SVM classifier [RFVL:4] 100 CAD System-II: (Normal, cirrhotic and HCC Liver): Design based on 2D-WPT multiresolution texture descriptors [RFVL: 10] 88.8 CAD System-III: (HCC and MET): Design based on texture features computed from IROIs and texture ratio features computed from IROIs and corresponding SROI. [RFVL: 9] 91.6 Note: MRS: Multiresolution scheme, RFVL: Reduced feature vector length, OCA: Overall classification accuracy, IROIs: Inside lesion regions of interest, SROIs: Surrounding lesion regions of interest. The CAD System-IV of Module 2 is designed to assist or provide second opinion to the radiologist for making differential diagnosis between FLLs using B-Mode US images. xxv The radiologists diagnose typical focal liver lesions (FLLs) easily by their classic sonographic appearances; however, the differential diagnosis between atypical FLLs from BMode ultrasound (US) is quite a challenging task faced in routine clinical practice, mainly due to existence of overlapping sonographic appearances even within individual classes of FLLs. Even then, B-Mode US is considered as preferred examination for characterization of FLLs, mainly due to its noninvasive, nonradioactive, inexpensive nature and real-time imaging capabilities. Therefore, a CAD system for classification of FLLs from B-Mode US images is highly desired. At the same time, it is worth mentioning that there are certain disadvantages associated with the use of B-Mode US for characterization of FLLs: (a) limited sensitivity for detection of small FLLs (< 2 cm) developed on cirrhotic liver which is already nodular and coarse-textured, (b) sonographic appearance of HCC and MET lesions are highly overlapping, (c) sonographic appearances of cystic metastasis and atypical cyst are often overlapping, (d) sonographic appearances of atypical HEM, sometimes mimic with atypical MET or HCC, and (e) difficulty to characterize isoechoic lesions with very slim difference in contrast between region inside the lesion and the surrounding liver parenchyma in some cases. Therefore, it is important to address these issues and design an efficient CAD system for FLLs using a comprehensive and representative image database with (a) typical and atypical cases of Cyst, hemangioma (HEM) and metastasis (MET), (b) small as well as large hepatocellular carcinoma (HCC), and (c) normal (NOR) liver tissue. As it is well known fact that US imaging has limited sensitivity for detection of SHCCs less than 2 cm in size, therefore in order to design a robust classification system, it is ensured that the constituent HCC images in the dataset offered a high degree of variability in terms of size and sonographic features. To ensure generality, the training data for designing the classifier was chosen carefully in consultation with experienced participating radiologists, so as to include all the typical and atypical image classes for Cyst, HCC, HEM and MET lesions as well as small and large HCC lesions for designing a robust classifier with representative cases for all image subclasses. Two sets of images were created for each image class, ROIs from one set of images were used for training and ROIs from the other set were used for testing to avoid any biasing. The Module 2 incorporates a CAD System-IV for classification between Normal, Cyst, HEM, HCC and MET liver image classes. In the present work, rigorous experiments were carried out for designing an efficient CAD system for characterization of FLLs. Radiologists xxvi visualize the texture patterns of the regions inside and outside of the lesion for differential diagnosis between FLLs using B-Mode US images. Accordingly, texture feature extraction from IROIs as well as SROIs was considered for the design of the proposed CAD system. Thus, IROIs (extracted from the region inside the lesion) and a corresponding SROI (extracted from the surrounding of each lesion) are inputted to Module 2 for classification between NOR, Cyst, HEM, HCC or MET liver image classes. The CAD system designs implemented in the present research work for characterization of FLLs using B-Mode US images, include designs (a) using PCA-kNN, PCA-PNN, PCA-NN and PCA-SVM based multiclass classifiers, (b) using hierarchical framework of PCA-kNN, PCA-PNN, PCA-NN and PCA-SVM based binary classifiers, and (c) using an ensemble of neural network classifiers. A brief comparision of classification performance obtained by different CAD System-IV designs implemented for the design of Module 2 of the proposed interactive system for diagnosis of liver diseases using B-Mode US images is depicted in Table 2. Table 2 Comparison of classification performance: CAD System-IV designs for Module 2 Experiment No. CAD System-IV Design OCA (%) 1 PCA-SVM based CAD system 87.2 2 PCA-NN based CAD system 87.7 3 PCA-PNN based CAD system 86.1 4 PCA-kNN based CAD system 85.0 5 PCA-SVM based Hierarchical CAD system 90.5 6 PCA-NN based Hierarchical CAD system 88.3 7 PCA-PNN based Hierarchical CAD system 91.6 8 PCA-kNN based Hierarchical CAD system 90.5 9 Hybrid Hierarchical CAD system 92.7 10 Neural Network Ensemble based CAD system 95.0 Note: OCA: Overall classification accuracy, NN: Neural network, PNN: Probabilistic neural network, kNN: k-nearest neighbour classifier. From Table 2, it can be observed that the neural network ensemble based CAD system yields the best performance for characterization of FLLs using B-Mode US images. Therefore, the NNE based CAD system should be used for the design of Module 2 of the proposed interactive system for diagnosis of liver diseases using B-Mode US images.