Brain Tumor MRI Image Segmentation Using Deep Learning Techniques

Brain Tumor MRI Image Segmentation Using Deep Learning Techniques offers a description of deep learning approaches used for the segmentation of brain tumors. The book demonstrates core concepts of deep learning algorithms by using diagrams, data tables and examples to illustrate brain tumor segmentation. After introducing basic concepts of deep learning-based brain tumor segmentation, sections cover techniques for modeling, segmentation and properties. A focus is placed on the application of different types of convolutional neural networks, like single path, multi path, fully convolutional network, cascade convolutional neural networks, Long Short-Term Memory - Recurrent Neural Network and Gated Recurrent Units, and more.

The book also highlights how the use of deep neural networks can address new questions and protocols, as well as improve upon existing challenges in brain tumor segmentation.

• Provides readers with an understanding of deep learning-based approaches in the field of brain tumor segmentation, including preprocessing techniques
• Integrates recent advancements in the field, including the transformation of low-resolution brain tumor images into super-resolution images using deep learning-based methods, single path Convolutional Neural Network based brain tumor segmentation, and much more
• Includes coverage of Long Short-Term Memory (LSTM) based Recurrent Neural Network (RNN), Gated Recurrent Units (GRU) based Recurrent Neural Network (RNN), Generative Adversarial Networks (GAN), Auto Encoder based brain tumor segmentation, and Ensemble deep learning Model based brain tumor segmentation
• Covers research Issues and the future of deep learning-based brain tumor segmentation

• Language: English
• Publisher: Academic Press
• Release date: Nov 27, 2021
• ISBN: 9780323983952

Book preview

Chapter 1

Brain MRI segmentation using deep learning: background study and challenges

Jyotismita Chaki

School of Computer Science and Engineering, VIT University, Vellore, Tamil Nadu, India

Abstract

The segmentation of brain tumors is an important component in medical image analysis. Early detection of brain tumors improves diagnostic methods and increases patients' chances of survival. Manual segmentation of brain tumors from large amounts of magnetic resonance imaging (MRI) images produced in clinical routine is a complicated and time-consuming process. Thus, an automatic way to brain tumor image segmentation is needed. This chapter aims to deliver an overview of deep learning (DL) based brain MRI segmentation. Automatic segmentation utilizing DL methods has recently gained popularity as these methods produce the state-of-the-art performance and can solve this issue better than the traditional. There are a variety of review papers available that concentrate on conventional approaches for MRI-based brain tumor image segmentation. Unlike others, the emphasis in this chapter is on the current developments of DL methods in this area, as well as the challenges and future aspects. First, an overview of brain tumors and segmentation methods is given. The state-of-the-art architecture is then addressed, with an emphasis on the current developments of DL methods. Finally, an overview of the challenges is discussed in the field of MRI-based brain tumor segmentation approaches into everyday clinical practice.

Keywords

Magnetic Resonance Imaging (MRI); Brain MRI segmentation; Deep learning; Convolutional Neural Network

1.1 Brain tumor and magnetic resonance imaging

The development of an abnormal cluster of cells within or near the brain results in the formation of a brain tumor. The abnormal cells disrupt brain function and affect a patient's health. The primary focus of study for the physician, clinical experts, and radiologist is brain imaging examination, identification, and treatment using accepted medical imaging methods. Brain image processing is deemed necessary because disorders of the brain such as brain tumors are dangerous and account for a significant number of deaths in different countries [1].

Brain tumors are primarily categorized into two types: malignant tumors (cancerous tumors) and benign tumors (noncancerous tumors). The World Health Organization categorizes malignant tumors into Grades I through IV [2]. Grade I, Grade II, Grade III, and Grade IV tumors are called pilocytic astrocytoma, low-grade astrocytoma, anaplastic astrocytoma, and glioblastoma, respectively. Grade-I and Grade-II tumors are less dangerous semimalignant tumors. Grades III and IV are malignant tumors that harm the patient's health and can result in death. Fig. 1.1 illustrates the pictorial representation of four different astrocytomas [3].

Figure 1.1 Pictorial representation of four different astrocytomas.

Single-photon emission computed tomography, computed tomography, magnetic resonance spectroscopy (MRS), positron emission tomography, and magnetic resonance imaging (MRI) are all medical imaging approaches utilized to deliver useful information about the size, shape, position, and brain tumor metabolism to help in detection. Although these techniques are utilized in grouping to deliver the most accurate brain tumor information, MRI is considered the most appropriate approach because of its good contrast of soft tissue and widespread accessibility. For the study of brain tumors using MRI, a variety of advanced methods such as diffusion tensor imaging, Perfusion magnetic resonance (MR), and MRS are generally used. MRI is a nonharmful clinical imaging method that utilizes signals of radio frequency to arouse target tissues, causing them to generate the internal images underneath. These various MRI modalities generate various types of tissue contrast images, which provide useful structural information and allow for the segmentation and diagnosis of tumors and their subregions. T1-weighted MRI (T1), T2-weighted MRI (T2), T1-weighted MRI with gadolinium contrast enhancement (T1-Gd), and fluid attenuate inversion recovery (FLAIR) are the four standard MRI methods utilized for brain tumor segmentation [4]. Fig. 1.2 depicts the pictorial representation of T1, T2, T1-Gd, and FLAIR brain MRI images [5].

Figure 1.2 Pictorial representation of T1, T2, T1-Gd, and FLAIR brain MRI images.

Although MRI acquisition varies depending on the system, approximately 150 slices of 2-D images are generated to reflect the 3-D brain volume. Moreover, when the requisite standard modalities' slices are merged for analysis, the data become extremely crowded and complex. T1 images are typically utilized to identify healthy tissues, while T2 images are utilized to demarcate the edema region, which creates a bright area on the image. The bright area of the acquired contrast agent (gadolinium ions) in the active cell area of the tumor tissue in T1-Gd images helps to distinguish the tumor boundary easily. As cancerous cells do not interfere with the contrast agent, which allows them to be easily distinguished from the healthy cell area on the same series. The water molecules signal is blocked in FLAIR images, which assists in separating the edema area from the cerebrospinal fluid.

For the diagnosis of a brain tumor, several image-processing approaches and practices have been utilized. The fundamental stage in image-processing methods is segmentation, which is utilized to extract the contaminated area of brain tissue from MRIs. The segmentation of the tumor area is a critical role in the diagnosis of cancer, care, and treatment outcome assessment.

There are several issues in segmenting brain MRI tumor images [6]. These issues can be concise as follows: (1) The most common MRI issue stems from the nonstandard intensity ranges collected by various scanners. Due to differences in magnetic field strengths and collection procedures, the brain MRI intensity values for the same patient can vary from hospital to hospital, (2) the brain tumor itself has no static form or prior experience. Brain disorder can manifest in any part of the brain and any form. Furthermore, the gray value range of this disease may intersect with the healthy area's gray value range, causing brain tumor segmentation further difficult, (3) MRI contains white ` noise throughout the collection procedure which is nonnegligible (4) spatial intensity variations in every dimension often influence uniform organization. The bias field effect of MRI is responsible for this. The smoothed low-frequency signal of image strength is affected by the MRI bias. An offset field correction preprocessing phase is needed here, which typically raises the pixel intensity around the brain, and (5) large brain tumors can distort the internal brain structure, rendering certain treatments difficult to perform.

Brain tumor segmentation employs a wide range of semiautomatic and automatic segmentation strategies and processes.

As manual segmentation takes a long time, the creation of robust automated segmentation techniques that deliver efficient and unbiased segmentation has become an important and common research field in recent years which can handle the major issues caused in brain MRI segmentation as discussed above.

1.2 Methods for brain MRI segmentation

Brain tumor segmentation methods are categorized into three groups based on the level of user interaction needed, such as manual, semiautomatic, or fully automatic.

1.2.1 Manual segmentation methods

Manual segmentation [7] is the process of outlining the target tissue by hand. Manual segmentation is tedious and time-consuming, and it cannot keep up with the increasing segmentation demands. Furthermore, each person has a unique style of segmentation, which causes variations in segmentation performance. Though manual segmentation has several drawbacks, it has the maximum segmentation precision to date and is frequently utilized as the optimal segmentation result for automatic segmentation.

1.2.2 Semiautomated segmentation methods

Semiautomatic [7] approaches necessitate user interaction for three key reasons: initialization, interference, and evaluation. In most cases, initialization is done by identifying a region of interest comprising the estimated tumor region for the automatic method to handle. Preprocessing system parameters can also be tweaked to fit the input images. Aside from initialization, automated algorithms can be guided toward an ideal result during the phase by obtaining feedback and making changes in response. Additionally, if the consumer is dissatisfied with the results, he or she may alter or repeat the procedure.

Although semiautomatic brain tumor segmentation approaches take less time than manual processes and provide more accurate outcomes, they are still subject to inter- and intrauser uncertainty. As a result, the majority of existing brain tumor segmentation research is based on completely automated processes.

1.2.3 Fully automated segmentation methods

There is no need for user intervention in fully automated brain tumor segmentation methods [8]. To solve the segmentation issue, artificial intelligence and previous experiences are primarily combined. Deep learning (DL) techniques are widely used in this field.

1.3 Deep learning

DL [9] models are normally trained to handle necessary tasks using manually developed features derived from raw features or data obtained from other basic machine learning techniques. DL allows computers to process useful representations autonomously, straight from the given input data, avoiding this manual and complicated phase. The most popular DL methods are different versions of artificial neural networks (ANN). The key thing that DL approaches have in particular is their emphasis on automatically learning the feature vectors. This is the primary distinction between DL methods and more ``traditional'' feature learning techniques. Build new features and conducting a task are combined into a single problem and thereby enhanced concurrently during the training phase.

Convolutional neural networks (CNNs), [10] an efficient way to learn beneficial features of images, have sparked interest in DL in brain tumor segmentation. Until CNNs could be used effectively, these representations had to be developed by hand or generated by less effective feature learning techniques. When it became feasible to utilize features learned straightaway from data, many of the hand-crafted features of the image were usually abandoned because they proved to be almost useless when compared to image features discovered by CNNs. Based on how CNNs are built, there are some powerful expectations encoded in them, which helps people understand how they are so efficient. Following are some explanations of the CNN building