A Review on Traditional Machine Learning and Deep Learning Models for WBCs Classification in Blood Smear Images

In computer vision, traditional machine learning (TML) and deep learning (DL) methods have significantly contributed to the advancements of medical image analysis (MIA) by enhancing prediction accuracy, leading to appropriate planning and diagnosis. These methods substantially improved the diagnoses of automatic brain tumor and leukemia/blood cancer detection and can assist the hematologist and doctors by providing a second opinion. This review provides an in-depth analysis of available TML and DL techniques for MIA with a significant focus on leukocytes classification in blood smear images and other medical imaging domains, i.e., magnetic resonance imaging (MRI), CT images, X-ray, and ultrasounds. The proposed review’s main impact is to find the most suitable TML and DL techniques in MIA, especially for leukocyte classification in blood smear images. The advanced DL techniques, particularly the evolving convolutional neural networks-based models in the MIA domain, are deeply investigated in this review article. The related literature study reveals that mainstream TML methods are vastly applied to microscopic blood smear images for white blood cells (WBC) analysis. They provide valuable information to medical specialists and help diagnose various hematic diseases such as AIDS and blood cancer (Leukaemia). Based on WBC related literature study and its extensive analysis presented in this study, we derive future research directions for scientists and practitioners working in the MIA domain.


I. INTRODUCTION
Traditional machine learning (TML) and Deep learning (DL) techniques are widely used for various applications and are extensively applied in the medical image analysis (MIA) domain [1]. In modern healthcare systems, MIA is an essential attribute, assisting medical experts wisely. MIA plays a vital role in diagnosing several diseases such as brain tumors, lung cancer, anemia, leukemia, and malaria. MIA processes various image modalities such as MRI, CT-Scan, Ultrasounds, Positron Emission Tomography (PET), Blood Smear images, and hybrid modalities [2]. In MIA, the image modalities play a vital role in detecting and classifying The associate editor coordinating the review of this manuscript and approving it for publication was Wai-keung Fung . hard and soft tissues of different body organs for diagnostic and research purposes [3]. MIA has dense contributions for computer vision experts in the investigated topic, where TML and DL play a significant role in leukocyte segmentation, cancer detection, classification, medical image annotation, and image retrieval in computer-aided diagnosis (CAD). The CAD and computer aided-detection (CADx) rely on effective TML and DL schemes because their performance directly affects clinical diagnosis and treatment process [4], [5]. It further assists the doctors in the diagnostic and treatment process, easing their traditional working mechanisms. The recent developments in information technology, such as high-speed computational resources, hardware design, and storage capabilities significantly impact CAD.  Formerly, key application areas of CAD system via TML and DL involve early-stage brain tumor detection in MR images and leukocytes analysis. It provides valuable information to medical experts, helping them diagnose different hematic problems such as AIDS and blood cancer (Leukaemia). The main aim of MIA is to assist medical experts, doctors, hematologists, pathologists, radiologists in the diagnostic and treatment process more effectively and efficiently. In the medical field, it has been perceived that the mainstream human body's diseases are recognized by analyzing leukocytes/WBCs [9]. The increase or decrease of leukocytes/WBCs and their morphological structure, such as size, shape, and color variations in blood smear images, indicate different human body disorders.
There are different types of blood cells, such as WBCs (leukocytes), RBCs (erythrocytes), and platelets (thrombocytes). Leucocytes are further divided into five subcategories: monocyte, lymphocyte, neutrophil, basophil, and eosinophil, as shown in Fig. 1. Various TML and DL techniques have emerged in the last two decades to segment and classify WBCs in microscopic blood smear images. Conventional techniques rely on manual analysis of WBCs in blood smear images, a time-consuming, challenging task, and prone to errors [6]- [9]. Automatic and CAD systems have a crucial role in clinical diagnosis and appropriate treatment [10]- [13].
Therefore, automatic analysis of WBCs in microscopic blood smear images is gaining popularity because it can decrease the workload on hematologists and provide quick, efficient, and accurate results to assist medical experts in the diagnostic process [14]. There are mainly two ways to achieve automated WBCs classification in blood smear images, i.e., TML and DL techniques, which have great potentials to develop such automatic systems that can make medical hematology more efficient [14]- [16]. The General overview of TML and DL Models for WBCs classification in blood smear images is shown in Fig. 2. Different CAD systems can automatically diagnose numerous hematic types, such as AIDS and blood cancer (Leukemia), by analyzing leucocytes [15]. In TML, there are interconnected steps involved, such as segmenting ROI and extracting features followed by optimal classification. A variety of TML techniques are available, i.e., manual, semi-automatic, and automatic segmentation techniques to segment ROI from an image [16]. Features extraction is another step in the TML approach. However, selecting an optimal feature extractor is challenging due to varying feature dynamics, such as geometric invariance and photometric invariance. Nowadays, the vast emergence of DL approaches has resulted in high-performance MIA models, especially in clinical hematology using blood smear images [18]- [33].
This research provides a comprehensive survey of the available TML and DL techniques and their medical imaging applications, mainly targeting WBCs classification in blood microscopic images. There have been several surveys on MIA using TML and DL techniques and future trends focusing on MRI, CT, X-rays, but microscopic blood smear is a rarely addressed problem [17], [18]. Therefore, this study intends to fill this gap by analyzing state-of-the-art TML and DL techniques for MIA, particularly leucocytes classification methods in blood smear images. The proposed research's primary focus is to provide a comprehensive review of the use of TML and DL in MIA.
In the proposed study, a novel categorization is employed to find the most common TML and DL methods that are reviewed in separate groups according to the research focus and employed technique. This research also helps identify future research directions by following TML and DL techniques to classify leucocytes in microscopic blood smear images. The followings are some of the significant contributions of the proposed review study: • The outlines of this paper investigate different applications and uses of TML and DL models in MIA.
• This This research study also aims to identify available machine learning techniques for leukocyte classification and analyze the extent of accuracy, applications, and MIA contributions.
• We address the key challenges and requirements of TML and DL models, followed by its future directions and solutions for future research in MIA. The remaining paper is structured as follows; Section II describes the review methodology and papers scrutinization process in detail. Section III gives a brief introduction about MIA. In Section IV, we present the detailed summary and applications of the artificial neural network and leucocytes classification in microscopic blood smear images. In Section V, the current challenges and requirements are discussed. Future directions of the proposed review study are described in section VI. In the last section VII, we discuss about recent advancements in DL models, followed by conclusions of the proposed review work.

II. REVIEW METHODOLOGY
This section provides a detailed discussion about digital libraries used for conducting a formal research process in the proposed review study. A planned searching procedure is required to find the available literature that fulfills the searching criteria, to utilize the available digital resources purposefully [19]. In the proposed study, we incorporated both manual and automatic searches to get the most relevant research articles by following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) model [20], [21].
We performed both manual and automatic searches to fetch the most relevant content. Our searching strategy begins with an automatic search on electronic databases to retrieve relevant data followed by verification of results by MIA and leucocytes classification experts. In the proposed survey, we search for articles from the period of 2014 to 2020. All the included sources are searched automatically as well as manually using the predefined keywords. The above-mentioned keywords and string are checked on each database and its pattern is modified based on relevant results retrieval. Numerous keywords associated with the study's primary focus is based on the four research questions (RQ) that are designed keeping in view the Patient, Intervention, Comparison, Outcome (PICO) framework [22].
RQ -1: What are the different TML and DL techniques for leukocyte classification in blood smear images? RQ -2: What are the different applications of TML and DL techniques in medical analysis, especially leukocytes classification?
RQ -3: How are TML and DL techniques used in MIA, particularly for leukocyte classification in blood smear images?
RQ -4: What type of machine learning is practical and efficient for analyzing leukocytes in blood smear images?

A. RETRIEVED PAPERS SCRUTINIZATION CRITERION
The initially retrieved papers are subject to inclusion/exclusion criteria by following the PRISMA guidelines. Table 1 represents the inclusion and exclusion criteria to filter out irrelevant articles. The selection of research articles is completed in three steps. Firstly, duplicate papers are removed. Secondly, the paper title, abstract, and keywords are investigated for relevancy, and finally, the remaining research papers are included after a thorough investigation. The process of exclusion and inclusion criteria is applied to eliminate  conflict analysis and biasedness. A total of 1436 research papers are collected to review the literature based on the research focus during the article selection process. In the initial selection process, manual filtering is incorporated, and the papers are filtered using the relevant title, and 1106 papers are obtained. These 1106 articles are then filtered by observing the abstract and conclusion, finally leaving 922 papers. These papers are filtered by methodology and results in the next step, and 725 articles are obtained. Then the articles are filtered after reading the full contents, leaving 216 articles. We checked the remaining articles' quality by evaluating the methodology, full-proof results, journal's impact factor, and citations. After checking all these parameters, 80 papers are picked for the proposed study. After the completion of the paper's scrutinization process (paper inclusion and exclusion), the quality assessment is performed. Each research article is assessed against the scrutinization criteria. All research articles are reviewed, and the quality of the papers with respect to each research question is assessed. Each of the selected articles is read and analyzed manually by the authors. The publication channels used for the article searching and the stepwise selection process are presented in Fig. 3.

III. MIA
The process that can provide visual information of the human body to assist the radiologists and doctors in an efficient diagnostic and treatment is called medical imaging [23]. There are many image modalities upon which the doctors and medical experts rely for diagnosing diseases and prescribing treatment. These modalities include CT,  X-ray, MRI, microscopic blood smear images, PET, and ultrasound [17], [23], [24]. These imaging technologies play an essential role in MIA; doctors and medical experts can automatically detect and diagnose different chronic diseases by analyzing these images. They can also visualize different body organs for research [37]. The number of research papers explored in this field is shown in Fig. 4. The last two decades have witnessed extensive medical imaging usage in CAD, for instance, in applications such as for leucocytes segmentation and classification, tumor segmentation and classification, detection and classification of breast cancer, image-guided therapy, and medical image annotation [25]- [28]. It has, therefore, became an integral part of today's modern healthcare systems [29].

A. TML AND DL FOR LEUCOCYTES CLASSIFICATION IN BLOOD SMEAR IMAGES
The literature includes a sufficient number of recently published review articles on TML and DL techniques used in MIA. The most recent and relevant research works about TML and DL methods in medical imaging, particularly for the classification of leucocytes in blood smear images [30], are discussed in the subsequent sections. In the proposed study, the most relevant and recent studies are searched out using keywords ''leucocytes detection'' or ''leucocytes classification'' by filtering the recent papers. During searching, we found that there is an exponential research growth of using TML and DL methods for leukocytes analysis in blood smear images. Fig. 5. represents the overall research results of DL and TML techniques for MIA and its exponential growth in the last two decades.
TML approaches involve interconnected steps, i.e., image pre-processing, segmentation, feature extraction, feature selection, and classification. The pre-processing step includes image enhancement such as contrast adjustment, noise removal, and image sharpening. All these steps are applied to the input image before image segmentation [41]. There are numerous pre-processing techniques such as median filter, low pass filter, high pass filter, and Gabor filter. These are used normally for image contrast adjustment, image sharpening, and noise removal before image segmentation. TML has been addressed by several researchers for leucocytes detection and classification. However, accurate nuclei detection, separation of borders to recover overlapped cells, segmenting ROI, robust features extraction, and best features selection have become challenging and time-consuming using these approaches [31]- [33]. In this approach, after segmenting ROI, the next step is feature extraction. In traditional supervised learning techniques, the classification depends on choosing robust features descriptor and best features selection algorithm [31], which are the most crucial steps towards efficiency and accuracy of the adopted technique. The general overview of TML is shown in Fig. 6.

B. LEUKOCYTES CLASSIFICATION USING SVM
There are numerous supervised learning techniques available to deal with leucocyte classification, such as SVM, ANN, Naïve Bayesian, and Decision Trees. Hegde et al. [70] proposed a novel technique in which the authors first segmented the WBCs and then employed SVM to classify WBC cells into a normal or leukemic cell. Zhao et al. [69] proposed a novel technique to segment and classify Leukocytes in blood smear images. Color co-relation and morphological based segmentation are applied, followed by texture features extraction and classification using SVM to classify WBCs into its five subclasses [90]. Table 3 elaborates on the key contributions and applications of SVM for leucocytes classification in blood smear images.

C. ENSEMBLES, HYBRIDS, BAYESIAN, K-NN AND DECISION TREES FOR LEUKOCYTES CLASSIFICATION
In addition to ANNs and SVMs, which have significant contributions to MIA, hybrids, Bayesian, Ensembles, K-NN, and Tree models have also been applied to solve the problems in different sub-domains of medical imaging such as brain tumor detection, lung cancer detection, leukocytes classification, etc. Abdulkadir et al. [71] proposed a hybrid approach for WBC classification in blood smear images. Sajjad et al. [15] proposed a smartphone-based quality healthcare system for smart cities, in which an ensemble multi-class SVM is used to classify WBCs in blood smear images.    Tantikitti et al. [72] proposed a computer-aided diagnosing system to diagnose dengue fever disease. A multi-level threshold technique is used to segment leukocytes in blood smear images. This research has two decision tree models for classification. The first model was used to classify the type of white blood cells that are lymphocytes or Phagocytes. The second model is used to classify the dengue virus infection as positive or negative. In [73], a novel technique is proposed in which WBCs nucleus and cytoplasm are segmented using simple thresholding. After segmentation, some morphological operations are performed using ellipse curve fitting, followed by feature extraction. For feature selection, the sequential forward selection technique is incorporated, and finally, a naïve Bayes classifier is used to classify WBCs. Vogado et al. [74] used a hybrid approach for the classification and segmentation of leukocytes. In their proposed technique, CNN features are used as input to train the SVM classifier. A transfer learning is also utilized for further classification of leukocytes, as comprehensively given in Table 4.

IV. ANN FOR LEUCOCYTES CLASSIFICATION
ANN is a supervised learning technique inspired by the biological nervous system of the human brain. It involves input, output, and hidden layers that are linked together via weighted connections. The performance of any ANN technique depends on these weights, which are numerical values. The output layer generates results given the inputs based on weights, error function, and neurons in the hidden layer. Several research studies have applied ANN in the context of MIA due to its enormous applications, including leucocytes classification, brain tumor classification, breast cancer detection, and lung cancer detection. Some notable contributions and applications are summarized in Table 5.

A. LEUKOCYTES CLASSIFICATION BASED ON DEEP LEARNING
DL allows us to define a system in which the feature extraction is not designed by human engineers but learned from data using a general-purpose learning procedure [79]. In the field of MIA, deep learning achieved satisfactory performance and relatively easy to build an end-to-end network using CNN [80]. TML models are trained on manually extracted features, or they learn features via other simple machine learning techniques to perform different classification tasks. Therefore, DL techniques have attracted the researcher's attention and motivated them to explore DL's benefits for WBCs classification. Currently, DL has become a powerful research tool in artificial intelligence, speech analysis [81], natural language processing (NLP) [82], and medical imaging [83]. DL's use is also becoming an essential aspect as a pattern recognition tool in the field of MIA [84]- [86]. According to a recent review on DL based MIA [87], DL algorithms and particularly convolutional networks, have become a choice for many for analyzing medical data. These methods are particularly suitable to those areas where human-like intelligence is required to analyze large amounts of data. Additionally, good knowledge is needed to extract rich features from a massive raw data volume [88].
However, this task is challenging and time-consuming when a vast collection of data is to be handled efficiently. DL provides end-to-end learning and eliminates all extra overheads of selecting feature descriptors and feature selection, as shown in Fig.7. DL methods' significant advantage is learning and automatically extracting semantically rich features from the raw data [82]. This is the main difference between TML and DL models. DL's unmatched benefits have attracted a large research community and industries to use DL-based approaches for MIA. DL models can be classified into different categories such as convolutional neural networks [95], deep belief networks [96], Long short-term memory networks [97], Recurrent Neural Networks (RNN) [98], and deep autoencoders [99]. Convolutional neural networks (CNNs) is widely used in medical imaging [17].

B. LEUKOCYTES CLASSIFICATION USING CONVOLUTIONAL NEURAL NETWORKS (CNN)
CNN consists of multiple convolutional, pooling, and fully interconnected layers with activation functions. It is trained using gradient descent and backpropagation as any standard ANNs (see Fig. 8) [100]. Typical CNNs generally have a successive convolutional and pooling layer followed by a fully connected layer. A Softmax function is used at the output nodes to classify WBC's into its five respective categories, i.e., monocyte, lymphocyte, neutrophil, basophil, and eosinophil. Banik al. [101] proposed a novel CNN model for WBCs classification by fusing the features of first and last convolutional layers using the BCCD database. Choi et al. [102] proposed a CNN model with eight layers for WBCs classification. Karthikeyan et al. [103] presented an LSM-TIDC method to classify WBCs in blood smear images. Firstly, images are pre-processed, then texture and geometrical features are extracted using a multi-directional model. Finally, the extracted features are fed as a feature vector to deep convolutional networks for efficient and early detection of WBCs in blood smear images. In [14], the authors proposed a Regional-Based CNN using transfer learning approaches to classify WBCs in peripheral blood smear images. The overview of some recent articles using DL for leukocyte classification is shown in Table 6.

V. CURRENT CHALLENGES AND REQUIREMENTS
In this extensive literature review, we found the major research challenges and requirements, several key features, their applications, and advantages of TML and DL techniques for MIA, particularly for WBC's classification in blood smear images. In the last few years, there are certain standard and powerful TML and DL models developed for MIA, such as brain tumor localization and classification from MRI, leukocytes detection and classification in blood smear images, and lung cancer detection in CT images [124]. Still, there exist some significant challenges that the research community either has to accept or try to overcome. These challenges include the unavailability of publicly available large and good quality datasets, dedicated medical experts, and lightweight TML and DL techniques. Some of the challenges are related to the mathematical and theoretical underpinnings of many DML techniques [123], [124]. To overcome these challenges, unsupervised or semi-supervised systems are required [83]. The proficiency of semi-supervised and unsupervised methods in MIA will be compromised to avoid these issues. It is also challenging to move from supervised learning to unsupervised learning approaches without affecting the system's accuracy and efficiency. MIA applications and systems employing TML and DL methods are still far from perfect, leaving significant space for improvements.

A. UNAVAILABILITY OF PUBLICALLY AVAILABLE DATASETS
The major problem in the field of medical image analysis is the unavailability of publicly available datasets. To address this issue, the researchers need to encourage health organizations to make their medical data available; it can be interesting if quality data is publicly available for researchers. Moreover, initiatives that encourage open data from different health institutions worldwide are encouraged; some operation are also necessary (e.g., data from hospitals and conditional access to datasets). In all these cases, incentive mechanisms can be related to financial return, entertainment, or services to these institutions while providing quality data. The topic becomes more interesting for research when the data is available in massive amounts, just like other fields (e.g., video summarization [125], IoT [126], energy management [127], and so on.). It is vital to collect extensive and quality datasets with ground-truth labels for specific MIA applications. Moreover, such datasets can be used for benchmarking as well as hosting different competitions.

B. TRAINED PREDICTOR GENERALIZATION ABILITIES
The key issue with MIA and leucocytes detection and classification is to train a predictor. An ideal learning technique with a better balance of generalization ability and a computationally efficient heuristic model is required to overcome this problem. A learning paradigm that uses true or random labels and provides effective tools to deal with available datasets and efficient training algorithms are needed to train a model with remarkable generalization abilities. Learning with deep neural networks has enjoyed huge empirical success in recent years across a wide variety of tasks in the field of MIA, i.e., brain tumor detection, lung cancer, breast cancer detection, and leucocytes classification. Despite being a complex, non-convex optimization problem, simple methods such as stochastic gradient descent (SGD) can recover reasonable solutions that minimize the training error. More surprisingly, the networks learned this way exhibit good generalization abilities [128], even when the number of parameters is significantly larger than the amount of training data [129]. During model training, only minimizing the training error is not enough. Picking the wrong global minima can also lead to bad generalization behavior for the predictor. In such situations, generalization behavior depends implicitly on the algorithm used to minimize the training error. Different algorithmic choices for optimization, such as the initialization, update rules, learning rate, and stopping condition, will lead to different global minima with different generalization abilities.

C. TRUST-WORTHY METHODS TO BE FUNCTIONAL IN REAL-WORLD ENVIROMENTS
The existing TML and DL techniques are not good enough to be trusted without medical expertise to function in real world health diagnosis systems [130]. There must be an expert as well as technical skills to train a learning model for MIA and leucocytes classification. We need to explore such precise and trustworthy methods which do not need health experts and are implementable in real-world health applications.

VI. FUTURE RESEARCH DIRECTIONS
Considering the major challenges encountered by the MIA community outlined in section V, extensive work is demanded from the biomedical industry and research community to contribute to MAI and especially leukocytes analysis in blood smear images.

A. DATA AUGMENTATION TECHNIQUES TO FILL THE DATASETS DEFICIENCY
In this study, we have focused on the most frequently mentioned problem of unavailability of datasets in the field  of MIA and leucocytes classification. An extensive data augmentation technique and transfer leaning models are recommended to improve MIA and WBC's detection classification in blood smear images. There are several data augmentation techniques used to extend the existing data, i.e., classical image transformations like rotating, cropping, zooming, Gaussian blur, sharpening, edge detection, histogram-based methods, and finishing at Style Transfer and Generative Adversarial Networks.

B. MEDICAL EXPERTISE AND TECHNICAL SKILL ARE REQUIRED
In the future, computer-aided MIA-based diagnostic applications can benefit from the recent advances in TML and DL models. These models are already available on multiple open-source platforms such as Tensorflow, Caffe, and Keras [131]. However, selecting and training an appropriate machine learning model for a specific MAI problem is challenging due to limited medical expertise and clinical knowledge.

C. RESOURCE CONSCIOUS DL MODELS FOR LEUKOCYTES CLASSIFICATION
In recent developments, DL, i.e., GAN's (Generative Adversarial Networks), R-CNN, Fast R-CNN, faster R-CNN, and deep fusion of TML and DL techniques models have achieved higher performance in brain tumor detection, leukocytes classification, breast cancer detection, and other MIA tasks. However, their primary concerns are high computational cost and high memory requirements. So, computationally efficient and energy-friendly TML and DL models need to be explored for leukocytes analysis in blood smear images. Furthermore, such light weighted models can be easily implemented over resource-constrained devices.

D. END-TO-END LEUCOCYTES DETECTION AND CLASSIFICATION MODELS
Traditional learning techniques can be replaced by a deep neural network (DNN) based models. With the recent advancement of CNNs [132], end-to-end models are also gaining in popularity due to simplified model-building processes and the ability to classify leucocytes into its five categories. These models are based on data-driven learning methods and competition with complicated MIA models based on DNN. Different end-to-end architectures for leucocyte detection and classification in blood smear images, such as attention-based methods [133], [134] and CNN based model are also prominent.

E. UNIVERSAL EVALUATION FOR TML AND DL IN MIA
In MIA, the research community mainly relies on subjective evaluation techniques. However, this task is challenging, time-consuming, and can be prone to errors. Thus, further research is required to explore universal evaluation techniques that can automatically measure the performance of TML and DL models for MIA from different perspectives.

VII. DISCUSSION AND CONCLUSION
This study provided a comprehensive review of TML and DL techniques used for leukocyte classification in blood smear mages. We reviewed different TML and DL approaches to classify WBCs in blood smear images. The data are collected from primary studies published during 2014 to 2020. The current study's literature identifies 80 primary studies (articles published in journals, books, conferences, and online materials) defining TML and DL techniques for leucocytes classification in blood smear images and its applications in medical diagnosis. While reviewing the articles, we found that both TML and DL approaches have performed equally well with overall contributions in MIA. This study is focused on identifying different applications of TML and DL in MIA and leucocytes classification in blood smear images. The objective of this study is to gain insight into complex details of TML and DL by accumulating and analyzing the knowledge provided in the literature in order to facilitate further research in the field of MIA. This study shows that much work is still needed to investigate the use of TML and DL techniques for useful MIA and leucocytes classification in blood smear images. This study also aimed at identifying applications of advanced DL models other than leucocyte classification. However, it is found that almost all other medical diagnosis applications are either directly or indirectly related to TML and DL. The accumulation of all this information in this study will benefit the research community by identifying where they need to start in further research on TML and DL models for MIA.
In future these techniques will have tremendous contributions in the development of medical imaging, natural language processing and speech analysis. Beside WBCs, TML and DL techniques are also used for the detection and classification of different MIA domains i.e., MRI, CT, X-ray, Ultrasound images analysis. In the current study, we reviewed different TML and DL techniques such as SVM ANNs, Ensembles, Bayesians, neuro-fuzzy, hybrids, DL and CNNs which are used to analyzed blood smear image [15], [72]- [78]. In MIA, blood smear images are the emerging domain that achieved great attention by the research community since last three decades. Standard contributions and applications of TML and DL in MIA are presented in this study. Furthermore, we also identified the current challenges, future directions and solutions for the advancements of TML and DL models in the field of MIA and particularly for WBCs classification in blood smear images. In future, we aim to extend our survey by considering various MIA domains such as MRI, CT, Ultrasound, X-ray images by utilizing the potentials of TML and DL techniques.
TANVEER HUSSAIN (Student Member, IEEE) received the bachelor's degree (Hons.) in computer science from Islamia College University Peshawar, Peshawar, Pakistan, in 2017. He is currently pursuing the joint master's and Ph.D. degrees with Sejong University, Seoul, South Korea. He is currently serving as a Research Assistant with Intelligent Media Laboratory (IM Lab), Sejong University. His major research domains are features extraction (learned and low-level features), video analytics, image processing, pattern recognition, medical image analysis, multimedia data retrieval, deep learning for multimedia data understanding, single/multi-view video summarization, IoT, IIoT, and resource-constrained programming. He has filed/published several patents and articles in peer-reviewed journals and conferences in reputed venues, including IEEE TRANSACTIONS ON IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, IEEE  TRANSACTIONS ON INDUSTRIAL INFORMATICS, IEEE TRANSACTIONS ON INTELLIGENT  TRANSPORTATION SYSTEMS, IEEE INTERNET OF THINGS JOURNAL, IEEE  ALI SHARIQ IMRAN (Member, IEEE) received the master's degree in software engineering and computing from the National University of Science and Technology (NUST), Pakistan, in 2008, and the Ph.D. degree in computer science from the University of Oslo (UiO), Norway, in 2013. He is currently an Associate Professor with the Department of Computer Science, Norwegian University of Science and Technology (NTNU), Norway. He has over 65 peer-reviewed journals and conference publications to his name. He specializes in applied research focusing on deep learning technology and its application to signal processing, natural language processing, and the semantic Web. He is a member of the Norwegian Colour and Visual Computing Laboratory (Colourlab) and the IEEE Norway Section. He has served as a reviewer for many reputed journals over the years. VOLUME 9, 2021