Multiple ROI selection based focal liver lesion classification in ultrasound images
Highlights
► The classification of lesions depends heavily on the characteristics of the lesions. ► The characteristics are differently observed according to ROI selection methods. ► Existing ROI selection methods have limitation for guaranteeing robust classification. ► We propose novel multiple-ROI based focal liver lesion classification approach. ► The proposed approach attains enhanced and stable classification performance.
Introduction
Medical imaging is the technique and process used to create images of the human body for clinical purposes. Medical imaging modalities, such as positron emission tomography (PET), magnetic resonance imaging (MRI), computed tomography (CT), or Ultrasound (US), are widely used for non-invasive diagnosis of diseases. Among various medical imaging modalities, US imaging is the most widespread imaging modality for visualizing the human soft tissues, because PET, MRI and CT are of higher cost, having a lack of real-time imaging and greater operational inconvenience in comparison to US Mittal, Kumar, Saxena, Khandelwal, and Kalra (2011).
Liver imaging has been on one major application of the diagnostic ultrasound, which is helpful in early detection of liver diseases. Focal liver lesions (from diagnostic ultrasound images) pose a diagnostic challenge to radiologist and clinician alike. Unlike the diffused liver diseases, the focal diseases such as cysts and tumors are concentrated over a quite small area of the tissue and it is very difficult to identify the diseased part from the image alone, hence a biopsy test is needed to confirm the disease (Balasubramanian, Srinivasan, & Gurupatham, 2007). From a clinical point of view, an early non-invasive detection mechanism that identifies the different types of focal liver lesions from the digitized medical ultrasound image can play an important role in reducing the number of post biopsy hemorrhages-yielding patient emotional discomfort and financial expense.
Considerable research efforts have been done to develop effective methods for classifying focal liver lesions on ultrasound images. A typical focal liver lesion classification systems generally consists of three sequential steps: (1) extraction of region-of-interests (ROIs) containing a liver lesion from the whole ultrasound image, (2) feature extraction from the ROI in question, and (3) classification of the liver lesion based on extracted features. A lot of work has been done on feature extraction and classification for correctly classifying focal liver lesions on ultrasound images. However, little is published about proposing solutions to extracting ROIs including suspected focal liver lesions. Hence, to our best knowledge, developing effective ROI selection methods is still relatively new and challenging research issue to focal liver lesion classification application.
In the previous research, existing ROI selection methods can be divided into four methods. In this paper, we define these four methods as below,
- (1)
ROIin method which extracts a ROI within the focal liver lesion as shown in Fig. 1(a) and is a most commonly used method (Balasubramanian et al., 2007, Bommanna Raja et al., 2007a, Bommanna Raja et al., 2007b, Bommanna Raja et al., 2007c, Bommanna Raja et al., 2007d, Bommanna Raja et al., 2008, Bommanna Raja et al., 2010, Kim et al., 2009, Mittal et al., 2011, Poonguzhali and Ravindran, 2006, Poonguzhali and Ravindran, 2008, Poonguzhali et al., 2007, Yoshida et al., 2003, Zhang et al., 2010).
- (2)
ROIout method which extracts a ROI to include the boundary of focal liver lesion as shown in Fig. 1(b) (Abou zaid et al., 2006, Xian, 2010).
- (3)
ROIoverlaps method which extracts four ROIs located to overlap the half of each ROI with the ROIout as illustrated in Fig. 1(c) (Karule and Dudul, 2008, Karule and Dudul, 2009).
- (4)
ROIposterior method which extracts a ROI under focal liver lesion as shown in Fig. 1(d) (Kim et al., 2009).
The classification of focal liver lesions depends largely on how to represent the characteristics of lesions such as internal echo, morphology, edge, echogenicity, and posterior echo enhancement (Zhang et al., 2010). However, the main limitation of previously developed ROI selection solutions is that they may not be able to represent show all characteristics of liver lesions in a reliable way. Because the ability to observe each of the ultrasonic characteristics could differ depending on the type of extracted ROIs, as shown in Table 1. Firstly, the internal echo can be intensively observed with ROIin method, because it is referring to echoes from the inside of the lesion. As ROIout method only includes the boundary of a focal liver lesion, the morphology (which is referring to the form or structure of the lesion) is only observed with ROIout method. The edge refers to the border between the lesion and normal tissue, then ROIout and ROIoverlaps methods are related to detect the edge characteristic. In diseased states, the echogenicity of a lesion can be altered, either more echogenic (hyperechoic) or less echogenic (hypoechoic) than normal tissue. The ROIoverlaps method (which includes the normal tissue and lesion) is strongly related to observe the echogenicity. Finally, as the posterior echo enhancement means a variation of echo measured from tissue behind the lesion, the ROIposterior method can well represent this characteristic.
In order to classify focal liver lesions, it has been widely accepted that the characteristics of the lesions have to be well represented. To represent the characteristics of lesions, the previous researches have been focused on developing feature extraction methods. For improving the classification of extracted features, the study on using various classification algorithms has been performed. On the other hand, only a few studies have been performed so far. In this paper, we propose a new and novel multiple ROI based focal liver lesion classification approach. The proposed approach can achieve better and more stable classification performance of focal liver lesions by representing all characteristics of liver lesions. In the previous ROI selection methods, ROIoverlaps method looks similar with the proposed method. But, ROIoverlaps method could not represent all ultrasonic characteristics as shown in Table 1. To show the effectiveness of the proposed method, the proposed method is compared to existing ROI selection methods (including ROIoverlaps method) according to generally used features. Also, the proposed classification approach is compared to the existing focal liver lesion classification approaches (Bommanna Raja et al., 2008, Poonguzhali and Ravindran, 2006, Xian, 2010). To verify the usefulness of the proposed approach, experiments were performed using 150 ultrasound images containing 150 liver lesions (50 cysts, 50 hemangiomas and 50 malignant lesions).
The rest of this paper is organized as follows. Section 2 gives an overview of the proposed multiple ROI based approach. Then, the effect of multiple ROI selection is analyzed based on aspect model in Section 3. The experimental results are provided in Section 4. Finally, conclusion is given in Section 5.
Section snippets
Overview of proposed multiple ROI based focal liver lesion classification
In this section, we present an overview of the proposed multiple ROI based focal liver lesion classification approach that can make efficient use of ROIs (see Fig. 1 for examples) to attain the stable and better performance. Fig. 2 provides the overview of the proposed method. The proposed method largely consists of three sequential steps: (1) multiple ROI selection, (2) feature extraction, and (3) classification.
From an input ultrasound image I(x, y), we select firstly L ROIs each denoted by ROI
Analysis of multiple ROI selection based on the aspect model
The proposed multiple ROI based focal liver lesion classification approach has been developed underpinning a statistical model, the so-called the aspect model (Hofmann, Puzicha, & Jordan, 1999). The aspect model has originally been proposed by Saul and Pereira (1997) in the context of language modeling, where it is referred to as aggregate Markov model.
The aspect model is a latent variable model for co-occurrence data which associates an unobserved class variable denoted by cm ∈ {c1, … , cM} such
Experiments
In this section, we demonstrate that the classification performance for focal liver lesion could be significantly affected by ROI selection approaches and feature extraction methods adopted. Moreover, we evaluate that the proposed multiple ROI selection based approach can achieve improved classification performance compared to other state-of-the-art ROI selection methods. In addition, we demonstrate that the proposed approach is considerably robust regardless of which type of feature extraction
Conclusion
The classification of focal liver lesion is usually affected by detecting ultrasonic appearances of lesions, and these ultrasonic appearances are differently observed according to ROI selection methods. Focal liver lesion classification approaches, using existing ROI selection methods, have a limit for representing all ultrasonic appearances of focal liver lesions. Thus, the classification performance using existing ROI selection methods were low and variable according to used feature
Acknowledgment
This work was supported by the R&D program of MKE/KEIT (10033702, Ultra high speed Parallel Beamforming & Signal Processing).
References (24)
- et al.
Neural network based focal liver lesion diagnosis using ultrasound images
Computerized Medical Imaging and Graphics
(2011) An identification method of malignant and benign liver tumors from ultrasonography based on GLCM texture features and fuzzy SVM
Expert Systems with Applications
(2010)- et al.
Feature fusion of side face and gait for video-based human identification
Pattern Recognition
(2008) - Abou zaid, A. S., Fakhr, M. W., & Mohamed, A. F. A. (2006). Automatic diagnosis of liver diseases from ultrasound...
- Balasubramanian, D., Srinivasan, P., & Gurupatham, R. (2007). Automatic classification of focal lesions in ultrasound...
- Bommanna Raja, K., Madheswaran, M., & Thyagarajah, K. (2007a). Evaluation of tissue characteristic of kidney for...
- Bommanna Raja, K., Madheswaran, M., & Thyagarajah, K. (2007d). Analysis of ultrasound kidney images using content...
- et al.
Quantitative and qualitative evaluation of us kidney images for disorder classification using multi-scale differential features
ICGST_BIME Journal
(2007) - et al.
Ultrasound kidney image analysis for computerized disorder identification and classification using content descriptive power spectral features
Journal of Medical Systems
(2007) - et al.
A hybrid fuzzy-neural system for computer-aided diagnosis of ultrasound kidney images using prominent features
Journal of Medical Systems
(2008)
Texture pattern analysis of kidney tissues for disorder identification and classification using dominant Gabor wavelet
Machine Vision and Application
Gene selection for cancer classification using support vector machines
Machine Learning
Cited by (43)
A hybrid CAD system design for liver diseases using clinical and radiological data
2018, U-Healthcare Monitoring Systems: Volume 1: Design and ApplicationsComputer aided diagnostic system for ultrasound liver images: A systematic review
2017, OptikCitation Excerpt :For K-means categorization afterwards optimal features were utilized. Jeon et al. presented a technique to classify focal liver lesion based on multiple ROI, to obtain more reliable and better classification performance [79]. This technique can be utilized to classify focal liver ailment, for example Hem, Cyst, and Malignancies.
Assistive system based on nerve detection and needle navigation in ultrasound images for regional anesthesia
2016, Expert Systems with ApplicationsCitation Excerpt :ROI localization is an essential step for successful segmentation. Recent research has demonstrated the importance of the ROI detection approach (Chang et al., 2010; Jeon, Choi, Lee, & Ro, 2013; Singh et al., 2014). They have shown high scores for ROI classification for tissues such as liver, prostate or thyroid.
Automated stratification of liver disease in ultrasound: An online accurate feature classification paradigm
2016, Computer Methods and Programs in BiomedicineCitation Excerpt :For solving various classification, ANN applications along with BP learning algorithm is extensively used. Jeon et al. [54] compared the NN classifier to statistical classifiers analysis using tissue characterization parameters from liver images. Kalyan et al. [55] conducted a comparative study of different texture features extracted from liver ultrasonic images by employing Multilayer Perceptron (MLP) by finding the best classifier that distinguishes between abnormal and normal conditions of the liver disease such as such as “fatty liver”, “hepatomegaly” and “cirrhosis” conditions.
Content-based medical image retrieval using fractional Hartley transform with hybrid features
2024, Multimedia Tools and ApplicationsNeural Network Based CAD System for the Classification of Textures in Liver Ultrasound Images
2024, Lecture Notes in Electrical Engineering