A Brief Survey on Breast Cancer Diagnostic With Deep Learning Schemes Using Multi-Image Modalities

Patients with breast cancer are prone to serious health-related complications with higher mortality. The primary reason might be a misinterpretation of radiologists in recognizing suspicious lesions due to technical issues in imaging qualities and heterogeneous breast densities which increases the false-(positive and negative) ratio. Early intervention is significant in establishing an up-to-date prognosis process which can successfully mitigate complications of disease with higher recovery. The manual screening of breast abnormalities through traditional machine learning schemes misinterpret the inconsistent feature-extraction process which poses a problem, i.e., patients being called-back for biopsies to eliminates the suspicions. However, several deep learning-based methods have been developed for reliable breast cancer prognosis and classification but very few of them provided a comprehensive overview of lesions segmentation. This research focusses on providing benefits and risks of breast multi-imaging modalities, segmentation schemes, feature extraction, classification of breast abnormalities through state-of-the-art deep learning approaches. This research also explores various well-known databases using ”Breast Cancer” keyword to present a comprehensive survey on existing diagnostic schemes to open-up new research challenges for radiologists and researchers to intervene as early as possible to develop an efficient and reliable breast cancer prognosis system using prominent deep learning schemes.


I. INTRODUCTION
Cancer is the second death-causing disease that affects worldwide women. Cancer is a disorder range of the lethal cell if left untreated leads to indolent lesions and mortality [1], [2]. Abnormal cells are created as a result of a genetic mutation that grows out of control and becomes cancerous due to the changes in its deoxyribonucleic acid [3], [4]. Benign (a noncancerous tumor) does not invade neighboring tissue while malignant (cancerous tumor) spread in multiple body The associate editor coordinating the review of this manuscript and approving it for publication was Wentao Fan . functions via the lymphatic system and elicits nutrients from the body tissues [5], [6]. The most dominant cancer types are lymphoma, sarcoma, carcinoma, leukemia, and melanoma. Carcinomas is the most widely diagnosed form of cancers.
The breast tissues are comprised of various connective tissue, blood vessels, lymph nodes, and lymph vessels. (Figure 1a) shows the anatomy of the female breast. It often establishes, when the breast tissues grow abnormally and cell division is not controlled that results in the formation of a tumor. The developed tumor can be invasive or non-invasive which usually starts in milk ducts or the lobules [7], [8]. Invasive cancer may start in lymph nodes which spreads VOLUME 8, 2020 This in different organs using blood vessels but cancerous cells often remain separated from the tumor [9], [10]. Moreover, breast cancer is classified into various subtypes based on their morphology, shape, and structure [11], [12]. Early identification of breast cancer can assist in the prognosis process which can successfully mitigate serious complications of the disease with higher recovery [13], [14]. Various medical multi-imaging modalities such as digital mammography breast X-ray images (DMG), Ultrasound sonograms (ULS), magnetic-resonance-imaging (MRI), Biopsy (Histological images), and computerized thermography (CT) are exercised for breast cancer screening and classification. The auto-detection of lesions, lesions volume and its contour in mammography images is a prominent sign which is most significant in detecting the distorted edge of the malignant and smooth edge of benign tumor. (Figure 1b) demonstrates the benign and malignant masses in a digital mammogram [15]- [17]. It truly helps radiologist's in investigating malignancy and quickly analyzing the lesions to forbid avoidable biopsies. Initially, the radiologists analyze the images manually and final decisions are suggested after the mutual consensus of other experts. The availability of many radiologists at the same time in under-developed countries is a key issue. Moreover, the precise analysis of the multi-class images depends upon the experiences and domain knowledge of the radiologist.
Furthermore, the initial identification of breast cancer needs comprehensive monitoring of biochemical indicators and imaging modalities. CAD systems can serve as a second option to resolve breast cancer multi-classification issues. It can serve as an inexpensive, voluntarily accessible, speedy, and consistent source of early diagnosis of breast cancer. It can also assist the radiologists in diagnosing breast cancer abnormalities which can significantly decrease the mortality ratio from 30% to 70% [18].
Recently, various machine learning (ML), artificial intelligence (AI), and neural network schemes are exercised for image processing. The key achievement of the CAD system is to build an authentic and reliable system that can limit experimental oversights and can assist in separating benign and malignant lesions with higher accuracy. These systems are used to enhance image quality for human judgment and to automate the readability process of images for better understanding and interpretation. Currently, various articles on breast cancer detections, segmentation, and classification using ML and AI techniques have been published [18]- [20]. Most of the previous studies emphasized ML schemes using binary classification for the detection of certain cancer like lung cancer, brain cancer, skin cancer, stomach cancer, kidney cancer, and breast cancer.
Jaffar et al. [21] and Khan et al. [22] proposed a novel deeplearning-based model for breast cancer screening and classification using mammographic images. Qiu et al. [23] proposed a technique based on deep learning methods that classify the breast masses without lesions segmentation and feature selection. Samala et al. [24] performed breast cancer binary classification by reducing the computational complexities of all types of mammographic images. Nascimento et al. [25] extracted the morphological features from ULS images using binary classification. Youk et al. [26] proposed a new ULS technique named as Elastography to differentiate the benign and malignant lesions of breast cancer. The authors [27], [28] developed deep-learning-based techniques for suspicious ROI segmentation and classification using MRI modalities. Rasti et al. [29] developed a robust DL model for ROI segmentation and breast tumor classification using segmented DCE-MRI images. De Nazar et al. [30] proposed a model by selecting the variable value of the threshold for the segmentation of breast masses. Choi et al. [31] designed a CAD model to extracts the ROI before the breast cancer 165780 VOLUME 8, 2020 classification. The ROI extraction is the seclusion abnormal breast tissues from irrelevant regions that increase the accuracy and also the big number of images needed for training and testing. Casti et al. [32] used QDA-LDA model for auto-localization and classification of asymmetry ROI because it directly related to the accuracy of doctor's predicting and treatment Nahid et al. [33] proposed an approach that extracts ROI patches from HP images for the classification of invasive and noninvasive breast cancer by CNN. Bejnordi et al. [34] and Feng et al. [35] performed a biopsy to classify the breast WSIs into different categories through the deep-convolution neutral network and achieves the highest accuracy in binary-classification of cancerous slides. Punitha et al. [36] used the depigmentation technique to overcomes the merging of the neighbor region problems that almost have similar properties. Strange et al. [37] focused on the classification and distribution of microcalcification based on the topological model and morphological aspects.
The key objective of this review to assists the researchers in developing a novel and robust CAD tool which is computationally efficient and can help radiologist during the classification of breast abnormalities. This comprehensive review has exploited key research directions based on various multi-image modalities, image segmentation approaches, feature extraction techniques, types of DL and ML algorithms, and performance parameters used to evaluate the classification models. Statistical analysis of CAD systems considering different aspects is also highlighted through graphical and tabular representations. Following are the key research findings: As per literature, it is observed that there are huge variations in shapes of breast (abnormal) tissues, so the benchmarks can be taken off during the screening process. The micro-calcification morphology is another significant factor for defining ROI, which is based on the distance between each micro-calcification. A fixed-scale approach is based on the distance between individual calcification used for defining the micro-calcification cluster while the invariant-scale is a pixel-level novel approach that visualizes the various morphology aspects (i.e., calcification cluster shape, size, density, and distribution) to the radiologist. Furthermore, histogram-based methods and selection of optimal threshold is an efficient approach for the segmentation and classification of masses and calcification. From literature, it is also evident that none of a study has implemented this approach before. A novel CAD system needs to be developed based on this approach to classify the calcification and masses. A content-based image retrieval is a new approach based on mammogram indexing and ROI patches classification. From literature, it is found that none of a study used indexing on ROI patches to classify calcification and mass using a mammogram. However, indexing and ROI classification-based CAD system needs to be developed with the help of expert radiologist to get precise results. Furthermore, some challenges faced by DL algorithms for breast cancer diagnostics are related to ultrasound images because of its low signal-tonoise ratio (SNR) comparative to others. However, echogram is a new ULS imaging technology, which is much cheaper for breast screening. So, the development of a new DL algorithm is a significant task to break through the echogram image analysis. The CT or MRI image modalities are spatial 3D data which are very large in size and need higher computation resources. However, the design of light models is an interesting research direction for training and inferencing.
After this introduction section, the rest of this paper is organized systematically and is as follows. Section II presents the searching process, CAD system for the detection of breast cancer abnormalities (masses and calcification). Section III explains the breast cancer digital repositories. Section IV discusses briefly machine learning schemes for breast lesions diagnostics. Section V discusses briefly deep learning schemes for breast lesions diagnostics. We make a discussion on the open subjects and perspective issues in the research of breast cancer in section VI. Finally, research is concluded and highlights future research directions and challenges in section VII.

II. MATERIALS AND RESEARCH METHODOLOGY
This section explains the searching and selection criteria of articles relevant to breast tumor prognosis through deep learning techniques using different multi-image modalities. [38], [39].

A. SELECTION CRITERIA FOR RELEVANT STUDIES
The primary concern of this research is to find the answer to queries relevant to the classification of breast cancer through deep learning schemes using various multi-imaging modalities. The following queries are considered while designing this comprehensive study.
1) Types of imaging modalities recently used for breast cancer classification. 2) Types of the dataset (publicly and private) used to build deep learning classification models. 3) Types of DL and ML classifiers were recently used for breast cancer classification. 4) Challenges faced by the classifiers in accurately detecting masses. 5) Types of parameters used to evaluate breast cancer classifiers.
Based on above-mentioned research concerns, various well-known databases which include Science Direct, Web of Science, Scopus, MEDLINE via PubMed, Springer, IEEE Xplorer, Google Scholar, etc are explored using breast or breast cancer keyword in conjunction with mammography, ultrasound, MRI, CT scan, biopsy histopathological images for suspicious mass detection, diagnosis, processing, multi-classification through CAD, (ANN, CNN, KNN) DL schemes and (DT, NB, LR, SVM, RF, LDR) ML schemes, etc. Later, based on the scrutinizing and selection criteria mentioned above, 252 most relevant articles published between 2014 to 2019 are selected and its year-wise distribution is presented in (Figure 2a).

B. CAD SYSTEM FOR BREAST CANCER DIAGNOSES
Computer-aided-detection systems act as a second reader for the detection of lesions which may show the existence of breast cancer and radiologist made the final decisions [40]- [42]. The primary aim of the CAD system is to diagnose the suspicious area of the breast and mark the regions of interests which can be lesions. Zemmal et al. [43] and Saraswathi et al. [44] revealed that CAD detection systems have enhanced the accuracy of the radiologist for the detection of breast cancer. This section discusses the methodologies for the screening of breast cancer through CAD.

1) BREAST CANCER SCREENING
Primary concerns of breast cancer are unknown but it exhibits serious complications based on gender, age, and genetic history of patients [19]. Early detection of the breast tumor is treatable due to its small size and can improve the surveillance of patients. Moreover, the proficient judgment and assessment of breast cancer based on breast density help the physicians for the detection of masses and calcification as shown in (Figure 3). Breast cancer screening based on the appearances of cancer symptoms recommended by the American Cancer Society is presented in (Table 1).

C. CURRENT MEDICAL IMAGING MODALITIES
Breast screening refers to the use of medical multi-image modalities for the detection of breast abnormalities, so the early diagnosing can prevent cancer from proliferating [45]. A variety of image modalities uses for breast cancer detection, however, few key factors such as FPs, cost-effectiveness, and workflow used for performance evaluation of image modality shown in (Figure 4a) In this section, we highlight a few current multi-image modalities used for breast cancer prediction through the CAD system as declared in (Figure 4b). The comparative analysis, ins and outs of different multi-image modalities are presented in (Table 2).

1) DIGITAL MAMMOGRAPHIC IMAGE (DMG)
Mammography is a low dose x-ray image modality used to classify the subtle changes in breast tissues using the CAD system. The x-ray beam easily travels in fibro-glandular tissues of the dense breast to examine the mass and calcification that is a prominent sign of breast cancer [46], [47]. The contrast between the mass and calcification is almost very less, morphologically diverse, and extremely irregular which is difficult to diagnose clinically. However, DMG images used two ways for breast cancer screening. In the cranial-caudal view, the complete breast with all glandular tissues is examined. The fatty tissues are visible in the dark strip and the nipple is depicted in the contour. Moreover, mediolateral oblique is the 90-degrees projection in which breast tissues are outside breast quadrant where axilla can be imaged [20], [21], [48]. The advantages and limitations of DMG for breast cancer diagnosing are presented in (Table 2). DMG images are distributed into three types such as: 1.) Screen film mammography (SFM), has primary advantages because it is directly printed on a huge film. The authors Khanet al. [22] and Dhunge et al. [49] proposed a novel model for breast cancer screening, segmentation and classification (normal, abnormal, benign and malignant) using SFM. 2). Full-field digital mammogram (FFDM), most prominent technology has good image contrast, efficient processing, and easily available at publicly dataset. Carneiro et al. [50] design a model using FFDM to classify the segmented maps into normal and abnormal breast lesions. Qiu et al. [23] proposed a holistic approach to classifying the breast masses without lesion segmentation and feature selection. 3). Digital breast tomosynthesis (DBT), is the most advanced 3D technology that takes many breast images with different angles and integrates. Samala et al. [24] implements a model for breast mass classification by reducing the computation of all types of DMG, like DBT, SFM, and FFDM. Apart from the popularity of mammogram, in few cases, dense breast tissues remain invisible during the screening that leads to misinterpretation of the cancerous area that increases the FN ratio. In this case, the radiologist refers to the ULS, MRI, CT, or biopsy for better diagnosis.

2) ULTRASOUND (ULS)
ULS is a noninvasive modality for fast visualization and diagnosing of breast tissue. ULS uses the high-frequency sound-waves for intrinsic analysis of breast tissues including chest wall without radiation immersion like in DMG and MRI [28]. Doctors often refer to ULS tests to examine the noninvasive breast cancer (mass or cyst) and also used to find breast abnormalities such, swelling, pain, and breast infection [51], [52]. Abdel-Nasser et al. [53] found that ULS is the best choice for diagnosing of dense, fatty, and thick breast tissues instead of using DMG. The detail of the advantages and disadvantages using ULS image are discussed in (Table 2). Nascimento et al. [25] extracts the morphological features from ultrasound image and perform the binary classification through deep learning method. Youk et al. [26] proposed Elastography, a newly developed ULS technique to differentiate the benign and malignant lesions based on tissue stiffness or hardness.

3) MAGNETIC RESONANCE IMAGING (MRI)
MRI is a diagnostics image modality that uses strong radio waves and magnetic fields to capture the 3D image of breast tissue and display it in a clear view than DMG, ULS [28]. The ins and outs of MRI for diagnosing of breast cancer as in (Table 2). Doctors usually refer to MRI to get detail information when cancer has been diagnosed. MRI machine takes many breast images with different angles and integrates. However, MRI leads to breast biopsy after diagnosing the suspicious regions. Amit et al. [27] developed a robust deep-learning-based model for suspicious ROI segmentation and classification of breast tumors using MRI images. R. Rasti et al. [29] extracted features from segmented ROI and feed into ANN, CNN for detection, and multiclassification of breast cancer. Bevilacqua et al. [54] developed the DL model for ROI segmentation and classification using segmented DCE-MRI images.

4) HISTOPATHOLOGICAL IMAGES (BIOPSY)
Biopsy imaging is a breast tissue analysis approach for the screening of breast tumor, therefore, many researchers use HP images for precise classification [34], [35]. Soft tissues are often taken form suspicious areas and fixed on microscope slides. The stained microscopic slides are examined by a pathologist and changed into WSIs (digital-color-images). The experts used approaches to take out cells from a suspicious region are fine-needle aspiration (FNA), core needle biopsy, vacuum-assisted biopsy, and surgical biopsy (excisional or incisional). Nahid et al. [33] classify breast cancer into several categories instead of binary classification and WSI color images permit the creation of many ROI images that train the model. Shibusawa et al. [51] extract ROI patches from HP images to classify the invasive and noninvasive breast cancer. Auto-classification of breast tumor through HP images has many advantages over DMG, ULS and MRI as in (Table 2).

D. CAD FEATURES FOR MASS AND CALCIFICATION DETECTION
There are mainly three early signs of breast cancer visible during the screening of mammogram images that are discussed comprehensively in this section.

1) CALCIFICATION
The detection of calcification is the primary foci of screening and has led to the development of CAD system [67], [68]. Calcification found in two types: 1). In Macro-calcification, a larger deposit of calcium in breast depending on age is foreseen, which is often non-cancerous [69]. 2). In Microcalcification, there are small spots of calcium that appears in the form of cluster. In most of the cases, the number of calcium cluster are cancerous or malignancy. The microcalcification considered a malignant tumor and it seems like 3-5 clusters occupied in 1cm2 area [70], [71]. The benign and malignant micro-calcifications are shown in (Figure 1b). Strange et al. [37] developed the microcalcification topological structure at an invariant scale to classify benign and malignant calcification.

2) BREAST CANCER MASSES
Breast mass is a group of tissues occupied by lesions which are considered as a prominent sign of breast cancer. Mass could be malignant or benign based on a morphological structure like density, shape, and margin characteristics. The ROI segmentation process is based on the shape and size of masses [71]. Benign masses (cyst) are often found in oval, lobular, and round shapes having smooth boundaries while the malignant have irregular edges with ill-defined speculated margins. The radiologist may propose additional breast tests based on the size and shape of the masses. Abdel-Nasser et al. [72] proposed a mass segmentation technique based on mass size and shape to detect the ROI and breast abnormalities.

3) ARCHITECTURAL DISTORTION
Architectural distortion includes the heeling of the previous biopsy after injury, so it is very difficult to diagnose the injured area. It can be malignant or benign cancer and commonly considered the third mammographic signs of breast cancer. De La Rosa Toro et al. [73] proposed the architectural distortion model to classify breast cancer abnormalities.

4) SPICULATE LESIONS
A lesion with ill-defined margin appears in star-shaped which is mostly found in malignant cases. The speculated lesions are characterized as benign when the low-density spicules occur in the loose structure and low-density area. Zwiggelaar et al. [74] use the linear structure of the spiculated lesions to identify the breast abnormalities.

III. BREAST CANCER DIGITAL REPOSITORIES
This review presented exploit some knowledge based on multi-image modalities and other information of the same patient that help in the reduction of false-positive results using auto-system. In all datasets, DMG and MRI images are used widely while ULS, IRT, and microscopic are used limited for breast cancer prediction. From literature, it is found that mammography databases play a significant role in training, testing, and evaluation of DL schemes. The comparison of public datasets based on origin, image size, image views, image format, image mode are presented in (Table 3). Commonly publicly available datasets are BCDR [58], MIAS [59], DDSM [60], Banco Web, mini-MIAS [72], WBC [75], IRMA [76], INbreast [77], BICBH and BreakHis used extensively and their distribution declared in (Figure 2b). The public datasets present a mixture of normal, benign, and malignant annotated images and also extensive variability of patients cases. Many existing studies use publicly databases [28], [45], [58], [78], [79] however, a few authors uses private [51], [80], [81] which collected from research centre or hospitals. VOLUME 8, 2020

IV. MACHINE LEARNING MODELS FOR BREAST LESIONS DIAGNOSTICS
Mammography and fine needle-aspiration are well-known procedures used for breast lesions grading and infiltration of other body organs but both techniques suffer from false-(positive and negative) limitations due to misinterpretation by experts [82]. However, Machine Learning (ML) based on developing techniques helps in the intelligent automated identification of breast abnormalities, finding useful and hidden information that improves the tumor prognoses capability by reducing the diagnosis errors. Furthermore, ML-based techniques can provide judgment support to experts for an opportunity of the initial prognosis of breast tumors. Several machine learning techniques applied in the retrospective studies for the prediction of breast abnormalities, mass segmentation, and classification using pattern recognition [83], [84]. The most commonly used machine learning techniques discussed in this study are, Support Vector Machine  (Table 5).
The Support Vector Machine (SVM) is a supervised machine learning (ML) classifier that provides high Accuracy (Acc), Sensitivity (SN), and Specificity (SP) in breast tumor classification as compared to other algorithms. SVM has a prominent advantage in theoretical research in recent years. SVM classifier is widely used for breast cancer identificatin, ROI segmentation, feature extraction [30], [31], [46], [58], [59], [68], [75], [85]- [88]. In this study, 63 publications used SVM classifiers for breast cancer diagnoses that are 27% of the selected studies. Ul Haq et al. [89] designed the SVM based model to classify and segment the malignant and benign candidates by using feature selection mRMR (Minimal-Redundancy Maximal-Relevance) algorithm and Chisquar FS algorithm. The results reveal that SVM shows high performance due to suitable features segmentation (18 features) from the WBC database and obtained 99.71% accuracy. Wajid and Hussain et al. [90] explored the SVM model to distinguish between breast abnormalities (mass and calcification) using data preprocessing techniques such as CLAHE, Histogram equalization, and obtained maximum 99% accuracy. In many studies, support vector machine works with the association of other classification techniques such TSVM [43], PSO-SVM [59], Linear Proximal SVM, Lagrangian-SVM, Smooth-SVM, Finite Newton-SVM [75], Successive-Enhancement Learning-based weighted Support-Vector-Machine (SELwSVM) [91] to attain more promising results.
K-Nearest Neighbor is one of the top ten algorithms because of its intuitiveness, simplicity, effectiveness, and a supervised ML technique. It widely used in breast cancer detection, segmentation and classification [51], [68], [87], [92], [93]. KNN focus on the correctness of classification by calculating neighboring distance using K-NN values. This study consist of 21 KNN based publications which is 8% of the total reviewed studies. Ali et al. [94] employed the efficient tetrolet transform and KNN algorithm to classify breast cancer into cancerous and non-cancerous using mammograms. The tetrolet transforms decomposed the mammogram into sub-bands to extracts the energy-based features that serve as input to KNN classier which obtained 92% accuracy. Raghavendra et al. [87] proposed KNN method for automatic identification of breast lesion and Gabor wavelet for feature extractions from a mammogram which achieved 98.69% accuracy.
Random Forest (RF) is a kind of bagging integration method that used a decision tree as an individual learner [58], [95], [96]. It performs more precisely on the test set than a single algorithm and has a certain advantage in anti-noise capability on other algorithms. This study consist of 16 publications that used RF that is 6% of total reviewed studies. Wang et al. [97] proposed an improved randomforest-based rule extraction method (IFGRE) which derives the interpretable and precise extraction rule for breast tumor diagnoses using WDBC and WOBC dataset. The patient data are registered by the SEER program such as sex, birth date, tumor site, diagnostic stage, tumor morphology. The experimental results show that IFGRE obtained 99% accuracy. Abdel-Nasser et al. [53] presented the Random Forest algorithms to segment and discriminate of breast masses using ultrasound modality. GLCM feature, binary pattern, and ROI extracted using five methods which obtained 97.33% accuracy and 99% AUC.
Naive Bayes (NB) classifier is a probabilistic machine learning algorithm based on the Bayes's Theorem. Naive Bayes extensively used in sentiment predication and classification of breast masses [58], [87]. NB randomly measures the probabilities of pattern recognition for a specific task that computes the results based on some observation. This review consists of 11 NB based articles that are 4% of the reviewed studies. Chen et al. [98] proposes a selective Navies Bayes (SNB) algorithm for efficient selection of attributes to classify the breast masses and improved the mass prediction accuracy. Abdara et al. [82] proposed the nested Ensemble (SV BayesNet and SV-Naive BayesNet 3-Meta Classifier) techniques that used stacking and voting method for automatic identification of benign tumor from malignant lesions. The proposed model improves the performance of the prognostic system and obtained 98.07% accuracy.
Decision Tree (DT) classification aims to estimate the discrete value of the objective functions. DT is a supervised machine learning method that has high rule induction ability and fast breast cancer classification speed [68], [87], [95], [96]. DT classifies the big data to discover the potential value after training the DT function. DTN uses a tree for mapping the relationship between classification attributes and results. In this review, 12 articles used DT classifier which is 5% of selected studies Suresh et al. [99] presented a hybrid model for forecasting of misclassified malignant lesions based on decision tree algorithm and radial-based function that achieved overall 99% accuracy. Raghavendra et al. [87] developed a model to segment the breast masses automatically and used Gabor wavelet for feature extraction from breast images which achieved 96.52% accuracy.
Logistic Regression (LR) model is widely used for binary classifications and also in practical problems such as mass segmentation and classification of breast tumor [100], [101]. The LR function provides a universal and convenient method for the detection of breast cancer. This paper consists of 09 articles that used the LR model which is 4% of reviewed studies. Sultana and Jilani [102] used the LR model to predict the various classes of breast tumors by extracting hidden information pertaining to several types of attributes that achieved 97.18% accuracy and 0.99% AUC. Dhahri [103] developed ML techniques to discriminate the breast tumor based on fracture selection using WDBC datasets and achieved 92% accuracy.
Fuzzy method dividing the large numbers of objects into different classes based on their similarity. Classes are arranged based on similarities between objects in the same class. Different similarity criteria are based on the type of objects which include Euclidean distance and cosine similarity. Fuzzy-C Mean is one of the leading clustering methods which widely used for breast cancer classification [60], [61], [69], [104]. In this study, 12 publications used fuzzy (fuzzy c-mean) algorithms that are 5% of the total reviewed studies. Shrivastav et al. [105] designed a novel edge detection method based on a fuzzy rule base system. The proposed model consists of three steps, input variable (gradient vector), set rule (Gaussian and triangular membership-function), converting output (demulsification method) used to achieves the maximum accuracy. Mohammad and Al-Ani [106] developed Fuzzy based model for accurate segmentation of complicated medical images to get the required features and pattern which performs accurate segmentation by changing parameters.
Linear Discriminant Analysis (LDA) is a classic subspace-based and effective feature extraction method that is used to extract a feature vector from a suspicious breast region [68]. LDA is widely used for breast cancer prediction, face recognition, language processing, and other fields [32], [87], [107]. In this study, 14 articles applied the LDA classifier which is 6% of the selected studies. Mansour et al. [108] proposed a model BC-CAD that focus on tumor segmentation, ROI Localization, feature selection, characteristic extraction and grading of breast tumor using Histopathological images and achieved 96.70% accuracy.
AdaBoost cascading framework has a high reception and moderate error reception rate. In the cascading framework, front level classifiers are relatively simple in structure and use fewer features for classification [109]. AdaBoost filter the negative samples during sample discrimination, only positive samples are sent to the subsequent classifier for processing and negative samples are rejected directly. In this study, 6 articles used AdaBoost which is 4% of the reviewed literature. Zhen et al. [110] proposed a computational computer vision techniques deep-learning assisted efficient-Adaboost (DLA-EABA) algorithm to characterize the breast masses, feature selection and extraction using multi-image modalities, and achieved 97.2% high-level accuracy.
In this study, some other supervised and unsupervised ML algorithms were also reviewed for breast masses prediction and classification. De La Rosa et al. [111] explored the Multiple Instance Learning (MIL) to discriminate the masses and calcification by extracting texture features and achieved 91.10% AUC. Mahersia et al. [112] presented Bayesian regularization Back-Propagation networks technique for auto-detection of breast masses which achieved 97.08% accuracy. Singh and Urooj et al. [65] proposed Adaptive differential Evolution Wavelet-Ann (Ada-DeWNN) model for texture feature characterization and feature extraction from ROI in mammogram which achieved 99% accuracy. Ribeiro et al. [81] developed the Optimum-Path Forest (OPF algorithm) for the identification and classification of masses present in the suspicious regions of the breast which achieved 99% accuracy. Nascimento et al. [113] developed Polynomial Classification Algorithm to classify the breast image into cancerous and non-cancerous based on texture point derived from ROI which achieved 98% AUC. Wu et al. [114] proposed Artificial Immune System (AIS) to discriminate against the breast tumor using texture, morphological feature extracted from ROI which achieved 96.67% accuracy.

V. DEEP LEARNING FOR BREAST LESIONS DIAGNOSTICS
Recent developments in computational techniques, significant advancement in image-processing technology, and prevalence of DMG images have opened the opportunity to resolve the early diagnosing of breast abnormalities using DL schemes [23], [115], [116]. The existing ML approaches are imperfect for precise detection of breast densities; however, the DL approaches to deliver the auspicious development in mass segmentation to overcome the false-positive ratio (FPR). Cai et al. [117] used well-known DCNN schemes to overcome the limitations of the CAD system for DMG diagnosing. This section comprehensively elaborates the study of DL schemes and ( Figure 5) presents the analysis procedure of the CAD system for classification of breast abnormality.

A. IMAGE ACQUISITION
Many annotated images are required for the development of an efficient DL model which is difficult to meet practically; therefore manually splitting and labeling the picture, data-enhancement, preprocessing needs for implementations [118]. Initially, the capture mammogram is converted to a portable gray format that does not obliterate the data when it is compressed. The image itself has no label and semantics. It must be segmented and labeled manually before it is used. However, images collected in reality often have certain disadvantages that affect the quality of the feature extraction [119].

B. DEEP LEARNING IN IMAGE PRE-PROCESSING
With the rapid development of imaging technology, a massive number of medical images are available [64], [120]. VOLUME 8, 2020 Preprocessing of medical images is an important task before training DL schemes that include augmentation, ROI segmentation, resizing, image enhancements, cropping, noise removing, and feature reduction.
The distribution of CAD tasks in which a total 93 of studies (out 252) performed multi-preprocessing task for detection of breast lesions are presented in (Figure 9b).
Ahn et al. [78] and Duggento et al. [121] adopted augmentation to training the model and performing class-label predication. The massive number of medical images is needed for the training of DL schemes, the augmentation approaches such as geometric transforms, patch extraction, synthetic minority over-sampling, and noise addition used to increase the instances of images artificially for precise results. Choi et al. [31] developed a CAD system to extracts the ROI before the breast cancer classification. The ROI extraction is the seclusion abnormal breast tissues from irrelevant regions that increase the accuracy and also increase the number of images needed for training and testing. Casti et al. [32] used the QDA-LDA model for auto-localization and classification of asymmetry ROI because it directly related to the accuracy of the doctor's predicting and treatment. Wang et al. [107] used median filters for image smoothing to suppresses the noise and to reduce the edge blurring. Vijayarajeswari et al. [85] used mean filters to replace each pixel from its surrounding region. Peng et al. [64] used denoising filters to smooth the edges by denoise the edges of the image during the preprocessing.

C. DEEP LEARNING IN IMAGE SEGMENTATION
The breast segmentation involves removal of background region, pectoral muscles, labels, artifacts, and other defects add during image acquisition which disrupts the detection of breast abnormalities. The image segmentation split the digital mammography images into ROI patches to identify the breast abnormalities which is used for feature extraction as declared in (Figure 6a) and (Figure 6b). However, the bad segmentation leads to unprecise feature selection which provokes the wrong classification. The existing studies [59], [67] includes segmentation approaches such as thresholdbased, split and merge, edge-based and region growing based presented in (Table 4). The mass segmentation includes the breast area segmentation based on intensity, texture, shape features and abnormalities segmentation.
Andropova et al. [28] and Oliver et al. [122] performs work on mass segmentation and computed the mass features with human intervention. The commonly used filters for contrast enhancement and noise reduction are mean, median, histogram equalization (HE), morphological method and contrast limited adaptive histogram equalization (CLAHE) that increase the readability of medical image as declared in (Figure 6a). De Oliveira [123] proposed SVM based method for mass classification based on the mass shape feature, texture feature, and intensity because it is not always possible to extract the mass boundary clinically. Dhungel et al. [49] proposed a hybrid mass segmentation model comprises of DL module and TRW methods which decline the segmentation errors and provide precise results. Zhu et al. [124] proposed a multistage deep end-to-end learning model associated with FCN and CRF for the mass segmentation and ROI extraction in DMG.

1) REGION BASED SEGMENTATION APPROACH
Region-based segmentation approach based on pixel uniformity inhomogeneity region and similar features of the segmented area such as geometry, texture, and intensity. Liu and Zeng [86] used the region-growing approach for mass segmentation which required an initial seed point that merges all similar pixels within the mass region. The region-based segmentation formula presented in (Equation 1) and (Equation2). and Pratiwi et al. [120] proposed a model based on splitting techniques which divided the segmented image into sub-region until no further discrete regions are left. Jian et al. [125] proposed a merging segmentation approach that starts with seed point which increases the region by the accumulation of similar adjacent pixels. (Figure 7) shows the merging approach of the nearest pixel with the initial seed point. Punitha et al. [36] used the depigmentation technique to overcomes the merging of the neighbor region problems that VOLUME 8, 2020 almost have similar properties. (Figure 8) a, b, c, shows the splitting while d shows the similar merging region.

2) CONTOUR BASED SEGMENTATION APPROACH
In the Region-growing approach, an initial seed point is required that merges all similar pixels within the mass homogeneity region [62], [91]. The authors Diz et al. [58] and Wang et al. [107] proposed a method to detects the edges of the image which is nearest to the boundaries of segmented images. There are several edge-based image segmentation techniques such as GLCM, Logic filter, and Haralick descriptor. Peng et al. [64] perform the mass segmentation based on the CAD system to classify the breast lesions without considering the human intervention and initial seed point.

3) THRESHOLD BASED SEGMENTATION APPROACH
It is a simplest approach used for foreground and backgrounds segmentations as declare in (Equations 3). Usually, the histogram-based method needs an optimal threshold value for the segmentation of objects. De Nazaré et al. [30] proposed a model by selecting the variable value of the threshold for the segmentation of breast masses.

D. DEEP LEARNING IN IMAGE FEATURE EXTRACTION
Feature extraction is a classification step used for feature calculation of ROI with associated properties such as size, shape, homogeneity, and tissue density. The algorithm's performance is based on features extraction and also enhance the classifier complexity with increasing the features to be extracted. Mostly features are handcrafted based on radiologist experience however, auto-feature extraction of medical images is a common segment of CAD approach. Many feature selecting methods such as PCA, LDA, chi-square test, etc. used to reduce the redundancy and complexity of feature space. Similarly, the pathologists assign breast cancer grading by considering the cancerous features of the cell, nucleus size, nucleus shape, and cell division ratio [64]. The general comparison on feature extraction are presented in (Table 5).
Pak et al. [129] proposed a model that extracts histogram features to classify the lesions which include mean, standard deviation, skewness, entropy, and energy. It widely used some visual information in the image, such as color, shape, texture, and other information. Hamoud et al. [104] and Kallenberg et al. [130] extracted some textural features which include contrast, correlation, the sum of (average entropy, variance), entropy, homogeneity, maximum correlation coefficient, correlation, variance, inertia, inverse difference, entropy difference, variance difference. Sun et al. [131] used DLL techniques to extract geometric features such as size, circularity, sphericity, irregularity. Dhahbi et al. [92] used kinetics features to classify the lesions like uptake rate, wash-out rate, curve index, signal improve the ratio. Saraswathi and Srinivasan [44] used binary Object features to classify the calcification like are area, projection, centroid, thinness, and aspect ratio, perimeter, orientation. The overview of features extraction and features selections schemes explains in (Figure 5).

E. DEVELOPED MODELS FOR CLASSIFICATION OF BREAST ABNORMALITIES
Classification is a significant approach used for breast tissue detections and segmentations such as pectoral muscle, fibro-glandular tissue, and fatty tissue. The existing ML approaches are imperfect for precise classification of densities; however, the DL approaches to deliver the auspicious development in mass classification to overcome the false-positive ratio (FPR) [45]. However, the networks that work as feature extraction are trained with large datasets to perform data representation which refers to the classifier to achieve tasks. Ragab et al. [46] presented a novel DCNN approach for feature segmentation(Threshold and region-based) and ROI grading using digitized mammogram images. Ahn et al. [78] classified pre-segmented masses into dense and fatly tissue using the transfer learning-based CNN method. Mammograms divided into patches using the data augmentation technique and achieved a 0.96% correlation coefficient. Kooi et al. [80] developed a deep conventional neural-network for the identification of the suspicious regions based on tissue densities and mass segmentation using a mammogram. Gaussian derivative filters are used for feature extraction that achieved 94% AUC, 92% accuracy. The most commonly used deep learning techniques are discussed in this review are, Convolutional Neural Network (CNN), Deep learning techniques (DL), Artificial Neural-Network (ANN). Furthermore, the detailed contributions of each classifier are summarize in (Table 5).
Convolutional Neural Network (CNN) is a multi-layer supervised learning neural network extensively used for complex AI problems such as breast cancer prognoses, segmentation, classification, pattern recognition, and natural language processing [47], [78], [132], [133]. The structural design of CNN comes from the study of animal visual cortical cell activation. Training of Deep-CNN with a small amount of medical data using transfer learning and augmentation approaches is very challenging. Recently, CNN-based algorithms are developed by considering lesions description to enhance the abilities of experts to identify and precise analysis of the initial stage of breast cancer [36], [42], [116], [121], [124], [134]. The retrospective studies show that the CNN-based model has improved the accuracy and reduce the FPR for the detection of breast cancer. This review consist of 27 CNN-based publications applied for breast cancer prognosis which is out of 11% of selected studies. Benzebouchi et al. [135] present a novel deep-learning-based CNN model for detection, localization of calcification, and breast masses using DDSM dataset and show remarkable performances for reducing of false-positive detection and obtained 97.89% accuracy. CNN also used with conjunction of several classifier such as Feed forward neural network (FFCN) [36], Fully-Complex Valued Relaxation Neural-Networks (FCRN) [44], Deep-CNN [62], [117], Transfer Learning-based CNN [78], Convolutional sparse autoencode (CSAE) [130], for breast tumor segmentation and achieved the high accuracy, specificity and sensitivity.
Deep learning (DL) is a core data analysis technology and has become mature in the field of data mining. DL technology resulted the promising performance in breast cancer recognition and classification [23], [49], [79], [118], [136]. The auto-feature extraction of lesions has great realistic significance so the DL algorithm is more effective for breast cancer identification. Recently, several deep learning models have been broadly categorized such as stacked denoising autoencoders (SDAE) [98], Principal component analysis-network (PCA-Net) [137] for feature extraction and segmentation. VOLUME 8, 2020           Sun et al. [131] explored the deep learning-based model (SSL algorithms) to extract the GLCM and geometric features such as size, circularity, sphericity, irregularity of ROI, and yields the remarkable performance which obtained 82.4% accuracy. In this study, 20 deep learning-based publications are proposed which is out of 8% of the selected study.
Furthermore, a human brain is composed of billions of interconnected neurons that receive, transfer, process and responds to the information using a chemical reaction. Similarly, an artificial neural network (ANN) is composed of a set of artificial neurons inspired by the biological neural network. ANN achieves promising outcomes in the delineation of high-resolution multi-image modalities in breast cancer prediction, mass segmentation, localization, and classification [44], [95], [138], [139]. Murtaza et al. [140] developed a computationally cost-effective, ensemble breast cancer classification-network (EBRC-Net) for breast cancer diagnoses at initial stage using histopathology images (BreakHis dataset) and obtained 97.74% accuracy. Six ML classifiers applied for feature extraction, however, false predictions reduced using three misclassification reduction models that show remarkable results. This study consist of 32 ANN-based publications which is 14% of the total selective studies. In many articles, ANN works with the association of other classification techniques such Back Propagation-ANN [52], Fuzzy-feed forward backpropagation neural-network (ACFNN) [60], Radial-Basis Function Neural-Network (RBFNN) [120], Probabilistic Neural-Network (PNN) [87], GA optimized ANN [54], to achieve the accurate and efficient performance. From literature, it is investigated that the ANN and SVM are widely used for breast cancer classification as declared in (Figure 9a).

F. PERFORMANCE METRICS ANALYSIS
After successfully training the dataset, test images are served as input to the classifier to evaluate its performance. In breast cancer diagnosing, the malignant or abnormal lesions are positive class samples while the benign or normal is a negative class sample. The popular metrics used for breast cancer classification as declared in (Figure 10a)   Accuracy metric is a proportion to the number of correctly classified instances in both abnormal patient or true positives and normal patients or true negatives. The comparison between different metrics for classification of breast cancer is presented in (Figure 10b). The 34 (out of 252) articles only calculates the accuracy and achieved the results between 100% to 74.92% [68], [141]- [143].
Sensitivity presents correctly diagnosed positive instances which is positive. It means that how many breast cancer patients are accurately diagnosed with total abnormal patient [31], [62], [86]. 13 (out of 252) studies calculate only sensitivity and achieved the performance between 98% to 82.4%.

VI. DISCUSSIONS
Breast cancer is a fatal disease that increases the mortality rate in women. Early intervention and clinical management can improve the diagnostic process. Based on its seriousness, large numbers of articles have been published as shown in (Figure 9a). So, it is problematic to summarize all research work related to abnormality segmentation using DL techniques in a single article. However, this research delivers a holistic approach where we tried to summarize the available breast databases, preprocessing approaches, segmentation approaches, development of DL models, performance metrics, and state-of-the-art findings of breast cancer segmentation and classification. The contribution of DL models for breast cancer prediction helped doctors significantly by providing the second opinion for the establishment of the final decision, which enhanced the satisfaction and confidence of the patients. The scarcity of experts and doctors in under-developed countries was a key issue, however, the CAD-based diagnostic system provided timely feedback which helped in improving the diagnostic process with a declining mortality rate. However, as a result of a comprehension survey from the current literature, our study suggested that the mammographic image is the most effective and reliable tool used for early breast lesions prognoses. It attained more prominent attention in providing significant information for early diagnosing of breast abnormal tissues which are helpful for possible treatment arrangements. Mammogram has soft tissue contrast that helps the doctors in revealing the location and magnitude of breast tumor due to distinct absorption of low radiation rate between normal and abnormal tissues.
The primary motivation behind this study was to assist researchers and doctors in the development of a robust CAD system which is computationally efficient and reliable for early prognosis of breast abnormalities. But, there exist several serious complications and challenges during their clinical implementation. The details of the existing up-to-date work on breast abnormalities segmentation and classification is illustrates in (Table 5). This study also shows that recent ML approaches can be limited for a particular kind of breast density and cannot be generalized to a worldwide population. However, the DL method uses hybrid and semi-supervised approaches to extract significant information for the segmentation and classification of breast lesions. Besides, to limit human intervention, the development of a fully automated CAD system is significant for masses segmentation. From literature, it is also found that the automated DL method needs higher computation resources which makes it unfavorable in a practical environment. Although, all of the techniques need a massive amount of annotated images for training and validating outcomes. The statistics of medical imaging modalities per classifiers used for breast cancer classification are shown in (Figure 11a). The availability of labeled medical images with image level and pixel-level annotation is a key issue because the image annotations from experts and doctors is a complex, expensive and time-consuming task.
Furthermore, from literature, it is also found that the development of DL approaches from limited medical images is an open research challenge, however, DL techniques often use a data augmentation approach to enhance the database. Due to the complex structure of the female breast, the availability of different medical images of the same patient in the publicly available database is another research problem as shown in (Figure 9b).
Apart from the building of an automatic DL model, financial support for the management and construction of the medical database is another research challenge. Besides, confidentiality and copyright issues for the availability of medical images is also a complex and difficult process. Furthermore, this study summarizes the recommendations to enhance the performance of DL approaches in segmentation and classification of breast abnormalities using multi-images modalities and are as follow: • Usage of preprocessing techniques to improve the image contrast like CLAHE, OTSU filters.
• Usage of invariant-scale approach for defining of ROI.
• Usage of context-based approach for ROI patches classification.
• Usage of histogram-based approach for the selection of the optimal value of threshold by using simple peak information for image segmentation.
• Usage of image cropping and down-sampling for more precise computation.
• Usage of augmentation approach to enrich the database.
• Usage of multi-image modalities of the same patient to enhance the reliability of the model.
• Usage of 3D image modalities database, if available such as US-SWE. VOLUME 8, 2020 • Usage of context and patient information in a multi-imaging modality.
• Usage of the available well-labeled database.
• Usage of ELM, TL, classification approaches for obtaining a promising outcome.
• Usage of interpretability of model-layer data to extract features.
• Usage of appropriate validation techniques comparative to the available dataset.
• Usage of recent libraries for the implementation of DL approaches such as PyTorch, Caffe, TensorFlow, Keras, MatLab.

VII. CONCLUSION
In this study, we systematically compared the strengths, limitations, and performance of recent DL and ML schemes by analyzing medical multi-image modalities as shown in (Figure 11b). From this study, it is also evident that with the advancement of DL approaches the process of breast abnormalities segmentation and classification is improved which truly assisted radiologists and researchers. Researchers often prefer to use public databases rather than private because of the fact that public databases contain a huge amount of records and are comprised of a mixture of normal, benign, and malignant cases. In addition, they also provide multi-modalities of images of the same patients.
Preprocessing is one of the most important steps that include augmentation, ROI segmentation, resizing, noise removing, image enhancements, and cropping. It is executed to remove irregularities in the images before the establishment of a training process for a DL scheme. Moreover, this research comprehensively examines the benefits and risks of existing literature for the development of a robust and reliable CAD system to limit the computational and time complexities related to the development of breast cancer diagnostic system. ML approaches are found imperfect for precise segmentation of densities; however, DL approaches helped in minimizing false-positive ratio (FPR) in the segmentation of masses. DL approaches need a huge amount of annotated images for training; therefore, to cope with the data scarcity issue, data augmentation is often adopted. Furthermore, this research also highlights significant research directions to appropriately select the DL technique, image-modality, and database for the segmentation of breast mass and calcification.