TEXTURE CHARACTERIZATION OF STROKE LESIONS IN NON-CONTRAST COMPUTED TOMOGRAPHY IMAGES OF NIGERIAN PATIENTS

CHAPTER ONE

INTRODUCTION

BACKGROUND OF THE STUDY: Medical imaging is a rapidly developing branch of modern medicine. It has in the past few decades evolved into a highly sophisticated diagnostic tool. It has improved the study of human internal anatomy and to an extent physiology and detection of pathologies which were previously impossible. The field has kept pace with other rapidly developing branches of medicine. At this stage of its development, detection of lesions and their interpretation is becoming an automated computer-aided process. It can now be safely said that machine vision has become an emerging part of radiology and imaging in medicine. Stoitsis et al. (2006) stated that advances in medical imaging technology and computer science have greatly enhanced the interpretation of medical images and contributed to early diagnosis. The bases for computer aided diagnosis (CAD) in radiology are medical image processing and artificial intelligence.

Cerebrovascular accident (CVA) or stroke, accounts for a significant proportion of neurological disorders seen in Nigerian hospitals (Ojini et al., 2003). Its incidence in Nigeria and other sub-Saharan African countries is on the increase (Myles et al., 2007). It carries high morbidity and mortality statistics in industrialized countries (Wolf et al., 1978; Gorelick, 1995; Sudlow and Warlow, 1996; Warlow, 1998) and Africa (Osuntokun, 1994). Stroke is the third most common cause of death worldwide after heart disease and cancer. It is reported to be the leading neurological cause of death in Africa (Howlett, 2012). The World Health Organization (WHO) (1988) defined stroke as a rapidly developing clinical syndrome of focal or global disturbance of cerebral function presumably of vascular origin, lasting longer than 24 hours unless interrupted by surgery or death. Other definitions of stroke have been derived from that of the World Health Organization. For instance, Holmes and Mirsa (2004), defined it as a focal neurological deficit lasting more than 24 hours and often preceded by transient ischaemic attack (TIA) in 10 - 15 per cent of the cases. Stroke occurs when blood supply to the brain is disturbed. This results in brain cells being starved of oxygen and consequently, some cells die while others are left damaged. Brain cells being permanent in nature, achieve only very limited recovery, leaving the patient with a permanent disability. A stroke may be due to infarction (ischaemic stroke) in 80 per cent of the cases or haemorrhage in the remaining 20 per cent. There are a variety of processes involving blood vessels which may lead to luminal compromise and cerebral ischaemia (Reeves and Swenson, 2008). Haemorrhagic stroke is usually associated with uncontrolled and longstanding hypertension. Clinically, ischaemic stroke presents as a focal neurological deficit of sudden onset, but there may be a step-like progression with headaches, complete loss of consciousness and vomiting as common signs and symptoms, unless the brainstem is involved (Holmes and Mirsa, 2004). Clinical presentation of haemorrhagic stroke varies according to the site, type and location of the bleed. Headaches, vomiting, focal neurological deficit and decreased level of consciousness are the characteristic signs and symptoms and there may be quick progression to coma (Holmes and Mirsa, 2004).

The clinical diagnosis of stroke and its subtyping is notoriously inaccurate (Chukwuonye et al., 2015; Sheta et al., 2012; Imarhiagbe and Ogbeide, 2011; Khan and Rehman, 2005). Neuroimaging is therefore essential for accurate diagnosis. Stroke remains one of the most important clinical diagnosis for which patients are referred to the radiology department for emergency imaging because timely and accurate diagnosis is critical in the management of patients (Mullins, 2006). Previous studies have highlighted the time-critical nature of ischaemic stroke diagnosis (Burnette et al., 1999; Jager, 2000; Tegos et al., 2000; Kidwell et al., 2000; Lev and Nichols, 2000; Keris et al., 2001; Burnette and Nesbit, 2001;). Ischaemic stroke has a narrow therapeutic window in the first few hours following stroke ictus and a dramatic rise in haemorrhage complications thereafter.

Non contrast head computed tomography (NCCT) has been suggested as the mainstay for early stroke diagnosis because CT scanners are more widely available in the communities and may be accessed much more easily (Mullins, 2006) than magnetic resonance imaging (MRI). This scenario is true of the Nigerian society and many other African countries because of the increasing utilization of CT. Computed tomography examinations are not only cheaper than MRI, they are also faster to perform. Thus, taking the time-critical nature of early stroke diagnosis into consideration, NCCT is the preferred first line imaging tool. Computed tomography is also considered to be very sensitive to early stroke and in most instances can provide information required to make decisions during emergencies (Chawla et al., 2009).

Computed tomography and other neuroimaging procedures will however not benefit the patient until the images have been accurately interpreted. For visual analysis and interpretation of stroke CT images, the radiologist seeks to identify affected areas of the brain by examining the dissimilarity between the left and right cerebral hemispheres. The challenges associated with visual interpretation of stroke CT images in Nigeria include the dearth of neuroradiologists (Atalabi et al., 2013) and the human errors of interpretation and diagnosis. Errors in visual interpretation result from poor technique, failures of perception, lack of knowledge and misjudgments (Robinson, 1997). Additional errors can come through the transmission of the radiologist’s visual impression, via non-visual means to the referring clinician (Sabih et al., 2010).

Visual interpretation can be improved upon by texture analysis which will make it possible for automated computer-aided approach to be used as a second opinion for clinicians especially in equivocal cases. Automatic method of stroke detection follows the same pattern as visual analysis and interpretation used by radiologists. Chawla et al. (2009) adopted this method in their study of automatic detection and classification of stroke from CT images of the brain. The method they used was based on the observation that the occurrence of stroke disturbs the natural contra-lateral symmetry of a CT slice. Accordingly, they were able to characterize stroke as a distortion between the two halves of the brain in terms of tissue density and texture distribution (Chawla et al., 2009). Some clinical applications of automatic detection and classification of stroke in CT images using texture analysis have been proposed notably by Bhat and Singh (2012) and Oliveira et al. (2009). Recently, Devi and Rajagopalan (2013) also proposed a method of segmenting stroke and non-stroke regions in magnetic resonance images and obtained encouraging results. Computer-aided diagnosis, also referred to as automated diagnosis, is not a common clinical application for stroke and other human diseases because there appear not be agreement on the best approach to it. The texture parameters and decision algorithms to be used are still subjects of debate. The present study is aimed at evaluating the relative accuracies of the four classes of statistical texture descriptors in automatic detection of stroke. The methods of automated diagnosis in medical imaging are anchored on texture analysis of medical images and artificial intelligence.

Artificial intelligence (IA) simulates the human brain or recreates it electronically. It is defined as the study and design of intelligent agents (Poole et al., 1998), where an intelligent agent is a system that perceives its environment and takes actions that maximize its chances of success (Luger and Stubblefield, 2004; Russel and Norvig, 2003; Poole et al., 1998). The simplest intelligent agents are programs written to solve specific problems. More complicated intelligent agents include human beings and organization of human beings such as a firm or a team. Artificial intelligence is based on the central characteristic of human beings; intelligence – the sapience of Homo sapiens. The human intelligence can be so precisely described that it can be simulated by a machine. The tools utilized in solving problems using artificial intelligence include search and optimization, logic, probabilistic methods for uncertain reasoning, classifiers and statistical learning methods and neural networks (Kapoor et al., 2015).

 

One very important stage in medical image processing leading to CAD is image texture analysis. Texture analysis of a medical image is the measurement of the quantitative parameters that constitute the image of a supposed lesion or normal tissue. This has the advantages of helping clinicians make accurate diagnosis and monitor disease processes under treatment. The analysis of texture parameters is a useful way of increasing the information obtained from medical images (Castellano et al., 2004). Four approaches are recognized in texture analysis and these are structural, model-based, transform and statistical approaches (Castellano et al., 2004; Materka and Strzelecki, 1998). The statistical approach to texture analysis is more popular with researchers involved in application of texture analysis in medicine because it yields an enormous amount of data and thus richer texture, so to speak. According to Ojala and Pietikäinen (2004), statistical methods analyze the spatial distribution of grey level values by computing local features at each point in the image and deriving a set of statistics from the distribution of local features. The reason behind this is the fact that the spatial distribution of grey values is one of the defining qualities of texture (Srinivasan and Shobha, 2008). The statistical method describes the spatial distribution of grey levels and their patterns. Statistical texture descriptors are made up of the grey level co-occurrence matrix, grey level run-length matrix, absolute gradient and histogram classes. Grey level co-occurrence matrix describes the relative positions of pairs of pixels with the same grey-level intensity while grey level run-length matrix describes the consecutive occurrence of pixels with the same grey-level intensity in particular directions. Absolute gradient calculates parameters related to variation of pixel grey-level values across an image, and the histogram represents the calculated grey-level distributions in an image. The different statistical approaches to texture analysis have advantages and disadvantages when applied to clinical situations. The co-occurrence matrix assesses texture on a pixel by pixel basis and allows for discrimination of images that are visually inseparable. It is computationally intense and gives rise to a large number of texture features. Therefore, there must be a mechanism to select the relevant features (Tuceryan and Jian, 1998). The co-occurrence matrix has been the most popular texture feature used for development of computer-aided detection and classification of lesions (Rajini and Bhavani, 2013; Zhang and Wang, 2007; Kabara et al., 2003). The run-length matrix is a third order histogram. It describes the number of consecutive pixels in a given direction having the same grey level intensity. Because the grey levels in the images are numerous calculating run-length matrix parameters may actually reduce texture information. Absolute gradient describes the spatial variation in grey level values across an image. It is a rather simple concept. It is usually used to emphasize contours or boundaries in images (Catellano et al., 2004). The histogram approach to texture is another simple concept. It is the count of pixels in an image that possess a given grey level value (Castellano et al., 2004). Many researchers and students are comfortable with the histogram method because of their familiarity with the histogram in lower level mathematics and statistics courses. In this study, the statistical approach to texture was adopted because the methods involved are versatile and also to compare the different statistical approaches to texture analysis.

The incidence of stroke in Nigeria and other sub-Saharan African countries is on the increase (Myles et al., 2007). Factors responsible for this increased incidence include change in diet, increase in cigarette smoking and alcohol consumption, inadequate exercise, increase in prevalence of obesity, and increase in other non-communicable diseases like hypertension and diabetes mellitus (Chukwuonye et al., 2013). In addition to the increasing incidence of stroke, the ratio of neuroradiologist to stroke patients in our locality is not encouraging. The aim of this study was to characterize and classify stroke lesions on non-contrast CT images using the four groups of statistical texture parameters. We also sought to cross-validate the classification with radiologist’s visual categorization identification of stroke lesions. This was to identify the most accurate statistical texture descriptor that may be used for computer aided diagnosis and classification of stroke lesions which would improve diagnosis and enhance patient management.

 

1.2 STATEMENT OF PROBLEM

There is increasing incidence of stroke in sub-Saharan Africa (Myles et al., 2007) which has resulted from increased incidence of its predisposing factors (Chukwuonye et al., 2013). Stroke poses a diagnostic challenge to clinicians practicing in our locality. Clinical diagnosis using the different weighted clinical scoring methods like Siriraj Stroke Score (Poungavarin et al., 1991; Chukwuonye et al., 2015; Rahman and Jamal, 2015;), Besson Stroke Score (Besson et al., 1995) and Allen Stroke Score (Nouira et al., 2009) is often inaccurate and thus radiological diagnosis is necessary. Computed tomography diagnosis is performed by the radiologist by visual inspection of CT images to identify the affected areas of the brain, the stroke type and quantify the extent of the lesion. The CT diagnosis is challenged by the dearth of neuroradiologists with experience in stroke detection and pitfalls due to human errors. Delays in obtaining fast, accurate and reliable diagnosis have led to high mortality from stroke, especially subarachnoid haemorrhage (Kowalski et al., 2004). The time-critical nature of stroke diagnosis therefore necessitates a simple, fast and reliable computer-aided automated process. There are no reliable data locally on texture parameters to construct algorithms for automatic detection and classification of stroke.

There has not been any holistic evaluation of the statistical approaches to texture analysis previously. There is no previous study aimed at producing an automatic detection and classification system for stroke lesions that included all the classes of statistical texture features for comparison of their performances in classification. It is therefore not clear which statistical texture descriptor is most suitable for computer-aided diagnosis of stroke.

The previously proposed methods lack generalization because they were based on data from a small number of patients and limited amount of data, focused mainly on ischaemic stroke and employed only one or two classes of texture features. With fewer participants in the studies and limited focus, it is possible that the classification algorithm may not have been comprehensive enough to detect all the possible cases of stroke.

1.3 OBJECTIVES OF THE STUDY

The main objective of the study was to characterize stroke lesions in NCCT of Nigerian patients using statistical texture parameters and identify the best statistical texture descriptor that will discriminate between stroke lesions and normal tissues as a preliminary stage in the development of an automatic detection system for stroke.

The specific objectives were to:

1. use statistical texture descriptors to analyze non-contrast computed tomography images of stroke lesions in Nigerian patients and identify the texture parameters that differentiate between normal brain tissue and stroke lesions using raw data analysis (RDA),

2. use artificial neural network (ANN) and k-nearest neighbour (k-NN) algorithms to classify brain tissue on non-contrast brain CT as normal, ischaemic and haemorrhagic using the obtained statistical texture features as the input data,

3. cross validate the classifications of brain tissue using ANN and k-NN with radiologists’ visual identification and categorization of stroke lesions and normal brain tissue using receiver operating characteristic (ROC) curve analysis,

4. to compare the accuracy of artificial neural network and k-nearest neighbour algorithms in classification of brain tissue of stroke patients, and

5. to determine the sex, age and anatomical distributions of stroke lesions in the patients.

1.4 SIGNIFICANCE OF THE STUDY

The result of this study will establish the best statistical texture descriptor that could be used in building automatic stroke identification and classification on NCCT of the brain. The result will also identify the algorithm more suitable for classification of brain tissue on NCCT of the brain. It will also provide the required database for computers to identify and classify lesions due to stroke, according to subtype on NCCT images which is hoped would improve diagnosis. The result will also serve as documentation for texture analysis of medical image research in general and CT images of stroke in particular which will be of immense benefit for future studies in the subject area especially in Nigeria.

1.5 SCOPE OF THE STUDY

The study was limited to patients who were clinically diagnosed to have had stroke and underwent NCCT examination of the brain in the CT suites of two private radiodiagnostic centres in Onitsha, Anambra State and Ibadan, Oyo State, Nigeria. The two radiodiagnostic centres received referrals from the teaching hospitals in both localities. The patients were referred from the teaching hospitals and peripheral private hospitals in both localities. The study also included chronic stroke patients who had follow-up NCCT of the brain following a recurrence. The data collection for the study lasted from May, 2013 to April, 2014. Other neuroimaging modalities relevant to stroke were not included in the study.