Computational Intelligence and Modelling Techniques for Disease Detection in Mammogram Images

Computational Intelligence and Modelling Techniques for Disease Detection in Mammogram Images comprehensively examines the wide range of AI-based mammogram analysis methods for medical applications. Beginning with an introductory overview of mammogram data analysis, the book covers the current technologies such as ultrasound, molecular breast imaging (MBI), magnetic resonance (MR), and Positron Emission mammography (PEM), as well as the recent advancements in 3D breast tomosynthesis and 4D mammogram. Deep learning models are presented in each chapter to show how they can assist in the efficient processing of breast images. The book also discusses hybrid intelligence approaches for early-stage detection and the use of machine learning classifiers for cancer detection, staging and density assessment in order to develop a proper treatment plan.

This book will not only aid computer scientists and medical practitioners in developing a real-time AI based mammogram analysis system, but also addresses the issues and challenges with the current processing methods which are not conducive for real-time applications.

• Presents novel ideas for AI based mammogram data analysis
• Discusses the roles deep learning and machine learning techniques play in efficient processing of mammogram images and in the accurate defining of different types of breast cancer
• Features dozens of real-world case studies from contributors across the globe

• Language: English
• Publisher: Academic Press
• Release date: Nov 16, 2023
• ISBN: 9780443140006

Book preview

Preface

Breast images are widely used to detect the different types of abnormalities associated with the breast. The medical sector is continuously exploring new ways to detect the different types of diseases with high accuracy. The improvements are visible in the scanning methods and the computational methods used for analyzing the scan images in the computer. Irrespective of the scanning methods, several efforts have been made to integrate Artificial Intelligence (AI)-based computational approaches to improve the success rate of the detection process. However, the scope for improvements still exists, which is the main motivation behind the contents of this book. This book provides novel technological concepts that can enhance the practical feasibility of the proposed subject. Apart from mammography, which is the main focus of this work, a few different modalities such as ultrasound imaging, molecular imaging, and histopathological image-based detection methods are also covered in this book. This adds more weight to the book. A wide spectrum of applications is covered in this book which will create an interest among biomedical researchers and computer scientists working in this field. This book, indeed, is a comprehensive product which will help readers to grasp the extensive point of view and the essence of the recent advances in this field. A brief introduction about each chapter follows.

Chapter 1 is an introduction chapter on the different scanning methodologies with special emphasis on mammography. It provides an overall idea on the current scenario in this field. The background information and the challenges associated in capturing high-resolution images are dealt in this chapter. The future trends of breast imaging are also discussed in this chapter. Chapter 2 deals with a brief introduction on AI approaches which are used to detect the diseases in breast images. The various difficulties associated with the current approaches are explored in detail. The future scope of AI-based disease detection in breast images is also dealt in detail. Chapter 3 deals with the Machine Learning (ML)-based approach for breast cancer detection. Random forest classifier is an ML method used in this work for disease detection. An exhaustive performance analysis is given to justify the efficiency of the proposed approach.

Chapter 4 deals with the mass detection process in mammogram images using Deep Learning (DL) approaches. YOLO is the prime DL-based method used in this work. The analysis is done in such a way to prove that DL-based approaches are efficient for the early detection of the diseases. Chapter 5 explores the possibility of application of pretrained DL models for disease detection in mammogram images. MobileNet and LSTM-based methods are used in this work for the experimental analysis. A comparative analysis is also given which validates the superior nature of the proposed approach. Breast image dimensions are usually a stumbling block to achieve results within a quick time. This problem is tackled in Chapter 6. An autoencoder-based approach is specifically framed to reduce the dimensions of the images without loss of accuracy. A basic classification approach is also implemented in this work to prove the efficiency of the autoencoder-based dimensionality reduction approach.

Chapter 7 deals with ML approaches and mastography images for the detection of cancer prognosis. Several ML approaches are used in this work to highlight the best possible method for efficient disease detection. Chapter 8 illustrates the process of microcalcification detection in breast images using ML approaches. Mammogram images along with decision trees and Support Vector Machines (SVMs) are used in this work. Convolutional Neural Network (CNN)-based implementation is also carried out for abnormality detection in breast images. Chapter 9 explores the application of pretrained models such as ResNet, AlexNet, etc. for cancer detection in mammogram images. The analysis is performed in a detailed manner to show the results of various stages of the implementation process.

An analysis of the different types of modalities for cancer detection in breast is given in Chapter 10. The pros and cons of the different modalities from the cancer detection perception are given in this chapter. The different stages of the cancer detection process are also dealt in detail. Deep Belief Network is the focus of chapter 11 in which ultrasound breast images are used for the experiments. An extensive analysis is also carried out to prove the need for DL approaches in breast cancer detection. Histopathological image-based disease detection in breast is the focal point of Chapter 12. Different transfer learning-based models are deployed to analyze the images for disease detection. ML model-based disease detection in breast images is carried out in Chapter 13. Histopathological images are used in this work. Chapter 14 deals with the fuzzy logic-based tumor detection in breast images. Mammogram images are used in this work for the experiments.

I am thankful to the contributors and reviewers for their excellent contributions to this book. My special thanks to Elsevier, especially to Ms. Carrie Bolger (Acquisition Editor), for the excellent collaboration. Finally, I would like to thank Ms. Emily Thomson who coordinated the entire proceedings. This edited book covers the fundamental concepts and application areas in detail. Being an interdisciplinary book, I hope it will be useful for both health professionals and computer scientists.

Dr. D Jude Hemanth

April, 2023

Chapter 1: Mammogram data analysis: Trends, challenges, and future directions

Karthikeyan Velayuthapandian¹,², Gopalakrishnan Karuppiah¹, Sridhar Raj Sankara Vadivel¹, and Dani Reagan Vivek Joseph¹     ¹Mepco Schlenk Engineering College (Autonomous), Sivakasi, Tamil Nadu, India     ²Department of ECE

Abstract

Mammography testing is one of the most important methods for the primary identification of breast cancer (BC), one of the most prevalent and fatal malignancies that mostly impact women globally. For precise medical diagnosis, multimodal image fusion offers a wide range of visual features. In the last decade, several articles have been reported describing methods for automatically detecting BC by examining mammograms. These methods were utilized to build computer systems that would aid radiologists and doctors in making more accurate diagnoses. Mammography can identify the microcalcifications that are present in several breast lesions. Radiologists can typically tell the difference between calcifications associated with benign illnesses and those associated with malignancy. Radiologists can use computer-aided detectiontechnologies to help them identify worrisome lesions in mammograms.

The number of deaths from BC can be minimized with early, efficient mammography identification and detection. Microcalcifications frequently influence the scope of surgical intervention, in addition to their importance in the timely and accurate diagnosis of BC via radiological assessment. Microcalcifications of a specific type are linked to poor biological and genetic features of the tumor and a poor prognosis. When compared to tumors comprising other forms of microcalcifications and noncalcified lesions, microcalcifications that are confined in the bigger ducts offer an independent negative prognostic sign. In this article, we will address the prognostic and diagnostic usefulness of microcalcifications found in the breast by mammography and provide an overview of the technical and theoretical framework for comprehending the clinical application.

Keywords

Brest cancer; Diagnosis; Mammography; Microcalcification

1. Introduction

Cancer is a collection of disorders in which the body’s cells cultivate out of control past their normal borders and attack various bodily regions. In accordance with the International Agency for Research on Cancer, there will likely be over 1.2 billion instances of cancer by the year 2050, making it the second biggest cause of mortality globally [1]. Based on the 2020 World Cancer Assessment by the International Agency for Research on Cancer (IARC), cancer is the leading or main mortality root for those aged 30–69 in 134 out of 183 nations. Males and females both die from lung cancer, but men are more likely to get prostate cancer, while women are more probable to develop breast cancer (BC). Agreeing to IARC, the tumor rate will increase from 14.1 million to 30.5 million between 2020 and 2050, while the amount of mortality will rise from 9.6 million to 16.4 million (Fig. 1.1). Cancer is brought on by a variety of intrinsic (like age-associated, gene vulnerability, hormones, inflammation, etc.) and extrinsic causes (like environmental influences, radioactivity, and way of life) [2]. Carcinogenesis, oncogenesis, or tumorigenesis is a multistage process that involves chronic proliferation, stimulation of oncogenes, mutation of tumor suppressor genes, tolerance to cell death, metastasis, and invasion immune cell denial, plus metabolic pathway remodeling [3,4].

Tumors can be categorized into various stages based on the anatomic severity of the condition. Tumor staging is crucial for prognosis, therapy planning, and treatment evaluation. One of the biggest obstacles to effective cancer therapy, even with the right tools for tumor categorization and staging, is early detection and tracking of the treatment’s effectiveness [6]. BC ranks as one of the general prevalent and dangerous tumors affecting females overall. Conferring to Fig. 1.2, BC accounts for 14% of all malignancies worldwide, is the most common disease among women, and has a high death and morbidity rate. It affects roughly 3 million women annually and raises the mortality rate for women. According to estimates, 685,000 women will pass away in BC in 2020.

Figure 1.1  Cancer registration history. Courtesy of WHO Global Burden of Disease, 2020; Department of Health, Hong Kong, China, 2020; Disease Registry, Macau, China, 2020.

Figure 1.2  Ten common cancer classes in the United States in 2022. Courtesy of Rebecca L, Siegel M, Kimberly D, Miller MPH, Ahmedin Jemal DVM. Cancer statistics. CA: A Cancer Journal for Clinicians 2017;67(27):7–30. https://doi.org/10.3322/caac.21708.

The BC spread stages in four major Indian smart cities are mentioned in Table 1.1.

Changing lifestyles, postponing weddings and having children, working long hours, and using hormone replacement therapy are major contributors to the rising occurrence of BC in industrialized nations [8,9]. The fundamental causes of high BC rates and deaths in developing countries include insufficient medical arrangement, inadequate mammogram screening, delayed identification, and scarce public-health education [10,11]. Fig. 1.3 illustrates the variety of treatments that are available to treat breast cancer [12,13].

Table 1.1

Source from Indian Medical Council for Medical Research, New Delhi.

On the contrary side, earlier tumor detection can increase the percentage of tumor patients who survive. As a result, there has been a trend toward the creation of technologies for the timely observation of tumors. The following is a comprehensive technological and theoretical background examination.

Figure 1.3  BC treatments.

1.1. Theoretical background

1.1.1. Sick lobe model

This idea of a sick lobe. Breast carcinoma is a lobar condition that begins early in an individual’s development when cells in one (or, extremely rarely, both) breast lobes gain early genetic defects that make this sick lobe more susceptible to the action of unpleasant stimuli. As a result, cells in the ill lobe of the mammary collect added genetic alterations than cells in the other, less vulnerable lobes of the same mammary. Over the period of a patient’s lifetime, the aggregation of abnormalities might cause the tumorigenesis of these units [14–17].

1.1.2. Neoductgenesis

Despite the fact that histological analysis of most carcinoma in situ (CIS) cases indicates an approximate number of ducts and terminal duct lobular units (TDLUs), the implicated ducts are significantly dilated and deformed due to the presence of tumor tissue and its byproducts. In contrast, some forms of CIS are characterized by an abnormally high density of duct-like patterns per unit surface area. The accumulation of tumor cells herniates the duct’s membrane to generate supplementary duct-like formations, and this phenomenon likely arises because of a blockage or disruption of the biological process of the alveolar flip, which would otherwise lead to the production of healthy TDLUs. Neoductgenesis is the term for this process, which typically includes the presence of calcifications [18–20].

1.2. Technical knowledge

Microcalcifications are so small that they cannot be seen with the human eye. Because of this, the surgeon needs radiological help at every stage of the diagnostic process for instances of calcification. To assist the physician in detaching nonpalpable tumors with acceptable restrictions, it is necessary to do preoperative radiological identification of the tumor using lead connectors or another method. The subsequent stage is specimen radiography, which directs the surgeon to take an appropriate sample from the tumor [21,22].

1.3. BC diagnosis using several imaging modalities

Screening decisions can be informed by risk assessments provided by breast cancer risk models. Multiple imaging techniques are being developed to detect this illness at its earliest stages. The identification and diagnosis of BC disorders rely heavily on medical pictures. Mammography, ultrasound, histopathology, magnetic resonance imaging (MRI), computed tomography (CT), positron emission tomography (PET), and thermography are the most common imaging tools used to diagnose BC nowadays. The various imaging techniques are all important in their own ways.

1.3.1. Mammography

When looking for BC in asymptomatic women, mammography is the bright mechanism. For initial and precise recognition of BC, nothing beats mammography. Breast X-rays are called mammograms. For breast screening and evaluation, it uses a small X-ray [23]. Mammography has done a much better job of finding microcalcifications and agglomerations of calcifications than any other method [24,25]. Screening images from mammograms have been shown to be technically superior, and hence they can be used in clinical assessment [18,26]. In addition, Digital Tomosynthesis Mammography, a modification of the original Mammography method, now offers 3D BC images. In a new development in mammography, an iodinated substance is infused intravenously during the mammogram itself, a process known as Contrast-Enhanced Digital Mammography (CEDM) (Fig. 1.4) [27].

1.3.2. Ultrasound

Images obtained using ultrasound, in contrast to those obtained through other forms of computed tomography, are often black and white and have a low image quality. In most cases, ultrasound pictures of malignant regions will have an uneven shape, a blurry appearance, and unclear margins. Mammography is unable to differentiate lesions from solid masses, although ultrasonography can effectively do so [28–30].

1.3.3. MRI

MRI is the most perceptive scheme presently accessible for diagnosing BC. MRI is able to clearly show the size, form, and position of BC lesions because it supports multiplanar scanning and 3D reconstruction methods [31]. MRI scans, however, are expensive and take a long time. In addition to mammography, MRI is the principal diagnosing technique for BC.

1.3.4. Histopathology

Histopathology is the benchmark for diagnosing BC, even above [32] other forms of medical radiography. They are necessary for the identification and therapy of tumor-related illnesses because of the phenotypic data they carry. Histopathology images (HI) have been used for BC multiclassification, although HI has important shortcomings in this area, including strong homogeneity of malignant cells, strong intraclass variation, and poor interclass distinction [33].

Figure 1.4  Mammogram timeline for BC diagnosis.

1.3.5. Thermography

Another imaging method for diagnosing BC is called mammary thermography, or thermal imaging, in which aberrant thermal patterns are strong indicators of mammary anomalies because of the greater amount of heat released by tumor tissue. Unlike other BC imaging modalities, breast thermography does not cause any discomfort to the patient and poses no risk to the thermographer, making it ideal for routine exams [34]. Thermography was being used alongside mammography to recognize BC at an initial level. Mammography is advised for teenage females despite the low dissimilarity of images of thick mammaries.

1.4. Risk factors

Gender and age are the most significant determinants of risk. Once the risk of BC has progressed to that point, one’s age is the key limitation. High age at birth yield, early menopause, history of BC in a parentage, the use of hormonal treatment, being tall, having a large body mass index after menopause, having dense breasts, drinking heavily, smoking tobacco, not getting enough exercise, being exposed to ionizing radiation, and having a short lactation time frame are all additional risk factors for developing BC.

1.5. Advantages in mammography

It is an original diagnostic imaging modality that utilizes a low-dose X-ray equipment to examine the mammary. Mammography is the broadly utilized BC tumor screening test. Mammograms are the outcome of the mammography technique. The mammogram is an X-ray imaging test that allows the medical professional to detect the anomaly. There are two distinct types of mammography: screening and diagnostic. Regardless of the onset of symptoms, screening mammography is used to examine the chest for any irregularities. According to medical knowledge, BC signs are only detectable after 2 years. Hence, consistent BC screening minimizes the likelihood of developing cancer. Diagnostic mammography is performed on women who have discovered anomalies such as masses, skin depressions, and nipple discharge.

In healthcare facilities, many mammograms are produced each day. As a result, radiologists must manually analyze a massive amount of images, which can take a long period to complete [35]. In reality, radiologists struggle to present a precise and consistent judgment for the vast majority of mammograms produced [36]. Additionally, their knowledge, expertise, and subjective norms all have a significant role in how accurate their assessments are. As a result, prediction accuracy tends to decline, making automatic mammography analysis and diagnosis increasingly desirable [37]. It has not been demonstrated that radiologists can effectively read mammograms visually. This happens because some physicians’ interpretations have been proven to be incorrect. According to research, radiologists typically make errors between 10% and 30% of the time, according to Lehman et al. [37]. With incorrect misperceptions, the rates of false negative (FN) and false positive (FP) results rise. In the report of Blanks et al. [38], reading mammograms twice improves the accuracy of the diagnosis.

Mammary biopsy procedures are necessary in combination with mammography because it might be challenging to extricate among malignant and benign tumors. In this sense, FP and FN misinterpretations may result in many consequences: a needless biopsy that results from a false positive error results in significant expenses and may be detrimental to patients. A true tumor may go undiscovered in a false negative finding, which could result in increased costs or possibly the patient’s mortality [39]. As a result, computer-assisted diagnosis (CAD) systems are crucial for assisting doctors in doing their duties with great efficiency and fewer mistakes [35]. CAD can assist in looking for worrisome symptoms and identifying tumors as benign or malignant. Microcalcifications and tiny tumors that are undetectable by self-examination are found through mammograms. The likelihood that a patient will receive effective therapy is greatly increased by early BC identification. Because mammography uses limited radiation, the risks are reduced.

The following are some of the most salient benefits of digital mammography.

• Breast mammograms improve the likelihood of detecting a cancer so small it could otherwise go undetected.

• Mammograms can find breast abnormalities before any symptoms appear.

• Because mammograms can detect BC at an initial level, without the need for invasive surgery, they are a valuable screening tool.

• Mammograms have been shown to have no adverse effects on patients.

This paper will follow the format below: BC and imaging techniques for diagnosing it are covered in the first section, followed by a brief discussion of the history of cancer screening strategies, detection schemes, and feature extraction mechanisms in the second. Sections 3–5 discuss various trends in mammogram data analysis as well as medical image analysis challenges. Following this will be a discussion of the principles and technology employed in the creation of point-of-care diagnostics as well as trends in clinical mammography practice. Future perspectives of mammography for improved clinical opportunities will be explored in the conclusion.

2. Related works

According to a review [39,40], healthcare professionals ignore 10%–40% of microcalcifications. Film and digital mammography are the two kinds of mammography, according to Gur [41] and Pisano et al. [42]. Film mammography creates the mammary using photographic film. In contrast, digital mammography digitally records the mammary, and the digitized mammogram is kept immediately in the system. According to Sampat et al. [43], irregularities in radiographs are identified by considering two indicators: mass and microcalcification. The goal of the techniques utilized for mass detection in mammograms is to locate and characterize the mass, if any, evident in the image. Form and contour are typical descriptors of the masses. Normal tissue has to be distinguished from masses using edge-based techniques. To differentiate among regular breast tissue and mass, various characteristics, including texture and shape, are retrieved from the targeted area.

2.1. Microcalcification detection

To automatically spot microcalcifications in mammograms, Oliver et al. [44] presented a boosting-based method. Employing local features collected from a filter bank [45] led to the development of the suggested methodology, which provides a description of microcalcification morphology. According to a novel quick proposed by Sankar and Thomas [46], mammograms may be modeled using a stochastic fractal coding scheme, and the presence of microcalcification, an early sign of breast cancer, can be identified. Using fractal encoding techniques [47], they contrasted the modeled mammography obtained with the baseline image, revealing the microcalcifications. Automatic detection of microcalcifications using a supervised method has been made possible due to the research of Torrent et al. [48]. The authors have detailed the suggested system’s morphological categorization, which is predicated on the retrieval of feature descriptors using a set of filters.

2.2. Classification of mass

Masses are a symptom of breast cancer. Regions of interest (ROI) on a mammogram might include both normal tissue and tumors that can lead to false positives. To solve the problems with massive categorization, Hussain et al. [49] compared and contrasted six alternative Gabor feature extraction strategies. Using mammography data of varying densities, Anand et al. [50] provide a technique for detecting structural deformities. The techniques for feature extraction were used to compute the ROIs from the Modified National Institute of Standards and Technology (MNIST) dataset. Automatic computer-aided identification and localization of mass in mammary images were pioneered by Isa and Ting [51]. Tissue structures revealed by mammography contrast fluctuations or image restoration were fostered.

2.3. Feature-based BC detection

Garge and Bapat [52] present an inexpensive adaptive filtering image processing solution implemented with the help of MATLAB software for detecting calcification on a somewhat difficult computational architecture. Microcalcification clusters in mammograms can be difficult to spot, but D'Elia et al. [53] and another study from 2004 describes complex methods for doing just that. Classification is achieved using the support vector machine (SVM) and orientation features (0,/4, 3/4, and/2), and Gabor filter alignment is assessed by measuring the detection accuracy [54]. A mean-centered region-based subdivision scheme, presented by Zaheeruddin and Singh [55], uses the position of the initial pixel to enhance the accuracy of the segmentation method as a function of the predefined threshold. For the purpose of producing the final classification, deep learning and machine learning methods were employed [56]. The experimental results showed that the whole fine-grained groups achieved equivalent precision and that the ensemble classification algorithm significantly improved the overall categorization compared to the coarse-grained