A review of microscopic analysis of blood cells for disease detection with AI perspective

Why the study needed

Previously, manual methods of blood cells analysis were used by pathologists. This might cause error in disease prediction since manual methods are dependent on experience and skills of pathologists. Hence, it is proposed that an automated system of image processing be developed using different algorithms. A simplified, automated and cost effective system is required for detection of diseases. Thus, the above components explained are analyzed for knowing health indication of human being and thereby detecting abnormalities, if any. Although many researchers contributed in the study, there is a need to explore the research in many perspectives.

1. Segmentation of different blood components is still having some shortcomings, such as overlapping cells during the staining.

2. There are different parasitic components in blood cells those need to be identified. So that an existence of a particular malady could be highlighted.

3. There are many challenging diseases like leukemia that have many sub-types depending upon the cell morphology. To detect the exact type of leukemia is still challenging.

4. In medical imaging, the use of artificial intelligence will have a critical issue that it is used as a black box. Hence, it could not be considered full proof and trusted at all times. The technique known as explainable artificial intelligence is the need of study in relation to these analysis concepts.

Who it is intended for

It is always a critical and crucial job for diagnosing diseases in the medical field Since these decisions are further related to a patient’s life. To provide significant contributions in the current diagnostic system is to be intended in many ways. This area is to be studied by a variety of disciplinary and inter-disciplinary researchers. Following are details that will show to whom the study is intended for:

Some basic terminology related to blood

Composition of blood

Figure 2 shows the details of different blood components. Blood is made up of the following elements- erythrocytes, known as red blood cells (RBC), leukocytes, known as white blood cells (WBC), and platelets. These are combinedly present within the plasma membrane. Leukocytes are further classified into two subcategories called granulocytes which consist of neutrophils, eosinophil and basophils, and agranulocytes, consisting of lymphocytes and monocytes. Blood plasma is a mixture of proteins, enzymes, nutrients, wastes, hormones, and gases. Platelets are small fragments of bone marrow cells. The primary function of red blood cells is to carry oxygen from the lungs to different body organs. Carbon dioxide is carried back to the lungs, which will be exhaled afterward. RBC count is the measure of different diseases in the human body. A low RBC count means anemia and high means polycythemia. White blood cells protect the body against infection. The different components of blood are identified to know about the health of a human being. Microscopic images of blood smear are analyzed for different disease detection.

Different areas where microscopic blood analysis is used

Blood smear images are observed under a good quality microscope and are investigated to identify any contamination inside the body (Poostchi et al., 2018; Al-Tahhan et al., 2019). Several abnormalities could be recognized from these microscopic images. Though there are different laboratory analysis techniques of blood examination available, image processing and computer vision could play a sound and satisfactory role in identifying maladies, preferably it analyzes the morphological abnormality of different components of blood (Razzak, 2017; Loddo, Ruberto & Putzu, 2016). The following are various areas where image processing and computer vision could be utilized for blood cell analysis (Othman, Mohammed & Ali, 2017; Alom et al., 2018; Lavanya & Sushritha, 2017; Xia et al., 2019).

Leukemia detection

Harmful reasons, such as leukemia, seriously influence the body’s blood-forming tissues and lymphatic framework. In leukemia, the white blood cells created by bone marrow are anomalous. Figure 3 shows the distinction between normal and leukemia cells. It has two significant subtypes, acute leukemia, and chronic leukemia. Leukemia can further be classified into other following types, namely, acute lymphocytic (ALL), acute myelogenous (AML), chronic lymphocytic (CLL), and chronic myelogenous (CML). Recognition of these malignant growth cells is done manually by microscopic image analysis and requires a competent pathologist. An improved automated system of leukemia detection is designed based on image processing and machine learning techniques, which ends up being proficient when contrasted with manual detection (Negm, Hassan & Kandil, 2018; Chin Neoh et al., 2015; Singh & Kaur, 2017; Shafique et al., 2019; Putzu & Ruberto, 2013; Jha & Dutta, 2019; Moshavash, Danyali & Helfroush, 2018).

Anemia and sickle cell detection

Abatement in hemoglobin or absence of solid RBC prompts anemia. It can cause contamination of viral sicknesses and issues identified with relaxing. Anemia detection is mostly done by identifying sickle cells in the blood. These sickle cells have a typical crescent moon shape. These cells are recognized and categorized as ordinary cells and sickle cells via automated algorithms involving computer vision and machine learning (Bala & Doegar, 2015; Elsalamony, 2016; Alotaibi, 2016; Javidi et al., 2018).

Figure 4 indicates the different blood cells’ components through microscopic examination and Wright stained smear of normal blood. For certain neurological disorders diagnoses such as Alzheimer’s and Parkinson’s diseases, there are no exact criteria related to clinical ways. For improvements in these kinds of diagnoses emerging metabolomics serves as a powerful technique. This includes a study of novel biomarkers and biochemical pathways (Zhang, Sun & Wang, 2013). There is one more challenge to identify proteins from the complex biological network that interact with each other as well as with the cell’s environment. Pseudo-Guided Multi-Objective Genetic Algorithm (PGMOGA) proposed that reconstitutes pathways by assigning orientation weighted network edges (Iqbal & Halim, 2020). A gene encoder presented that incorporates two-stage feature selection with an unsupervised type for the classification of cancer samples (Al-Obeidat et al., 2020). There is a requirement of finding the correct DNA sequence to get the desired information about the genetic makeup of an organism. A hybrid approach presented utilized Restarting and Recentering Genetic Algorithm (RRGA) with integrated PALS (Halim, 2020). For working on the different datasets, there is a need to get a set of visualization metrics. These are used to quantify visualization techniques—an approach of visualization metrics based on effectiveness, expressiveness, readability, and interactivity. Evolutionalry algorithm (EA) is used here as a case study. This methodology can also be utilized further for other visualization techniques (Halim & Muhammad, 2017). There is also a requirement for the extraction of information from larger datasets. A popular Frequent Itemsets (FIs) mining is a task to find itemsets in a transactional database. The graph-based approach is used for the representation of a complete transactional database (Halim, Ali & Khan, 2019).

Paragraph

Out of these different diseases, leukemia is one of the most dangerous in its later stages (Deshpande, Gite & Aluvalu, 2020). It develops blasts cells in the blood, which later affect the generation of normal white blood cells. As the number of these blast cells increases, the body gets shortened of healthy cells, leading to frequent infections. Different types of leukemia define the way to treat it. So it is always necessary that the type of leukemia be detected with great accuracy. The morphological differences in the normal and leukemia cells are shown in Fig. 3.

Generalized methodology

A generalized methodology for microscopic blood cell analysis is shown in Fig. 4. It consists of different stages like image acquisition, image segmentation, feature extraction, and disease detection. The blood sample is taken from the patient by a trained pathologist. After that, a slide is prepared to form a blood smear. The same slide is observed under the excellent quality microscope that will give an image. This image is taken either by camera directly or from an adaptor connected to a microscope. This image is considered for further analysis. Acquired images may have some unwanted regions and overlapping of different components of blood. This image is enhanced by applying a suitable image enhancement technique. So that a good quality image is now available for the analysis. After pre-processing, separation of different blood components is done, including separation of RBC, WBC, plasma, and platelets. By considering the generalized characteristics of blood components, segmentation is done. This process will separate the region of interest for further classification. RBC, WBC and, other components are further classified into their respective sub-classes. This will help specify a particular sub-class image for extracting features for analysis of blood cells, and depending upon the analysis, detection of disease is done. After the segmentation, different features are extracted by considering different components of blood. Features include size, shape, color, count of different blood components like WBC and RBC counts. Analysis of these features will further detect the disease or count the cells. Depending upon different features extracted, the decision about the disease could be taken. To make the decisions, different classifiers could be designed.

Decision tree classifier

It falls under the supervised learning type. It is employed for regression as well as classification. It has roots, branches, nodes, and leaves.

Figure 5 shows the different components of the decision tree. The bottom and the upper part are termed roots. A node represents a feature in the dataset. A branch connects two nodes. Different decision tree learning algorithms are there. These include ID3 (Iterative Dicotomizer3), C4.5, C5.0, CART. These algorithms have different characteristics that decide their uses in a particular application.

In the ID3 algorithm, before pruning, trees are extended to maximum size. C4.5 does not require categorical features. It is useful in defining distinct attributes in the case of numerical features. C5.0 has a comparatively smaller rule set and also takes less space than C4.5. The CART classification and regression tree are similar to C4.5; the difference is, it does not calculate the ruleset, and it braces numerical target variables. This type generates a binary tree.

Decision trees may be affected by the problem of overfitting, which could be further analyzed and is taken care of. Regularization is the process to be adapted during the generation of decision trees for compensating for the overfitting issue.

Random forest

It has many decision trees, with each tree having different training sets. Hence this algorithm is more effective in solving classification problems. An important issue in this algorithm is the selection of a pruning method and branching criteria selection. Popular gain measurement techniques here are the Gain ratio and the Gini index. This algorithm is based on the number of trees to be developed and samples used by the particular node. Figure 6 shows the random forest algorithm.

K-Nearest Neighbours (KNN)

This classifier employs the nearest instances of training in the attribute space. Here according to the k value of neighbor, the new sample value of the class is decided. For getting the class of the new vector, the closest k samples from training data are selected. There are specific methods for calculating the distances according to the classified samples. These are illustrated as follows.

(16) $$Euclidean,\sqrt{\sum _{i=1}^k{(x_i-y_i)}^2}$$

(17) $$Manhattan,\sum _{i=1}^k\left|x_i-y_i\right|$$

(18) $$Minkowski,{\left(\sqrt{\sum _{i=1}^k{(\left|x_i-y_i\right|)}^q}\right)}^{{\displaystyle \frac{1}{q}}}$$

Logistic regression

Regression can be of either linear regression and logistic regressions. Linear regression is a supervised regression algorithm while logistic regression is a supervised classification algorithm. It is categorized into different types. These are binary, multinomial, and the last one is the ordinal. The first type of regression is the,

Binary logistic regression model

This is the simplest type where the dependent variable can have either 0 or 1 showing only two possible types. So here we have predictor variables multiple but the target variable is binary or binomial.

(19) $$h_{\theta }(x)=g({\theta }^{T}x),where0\le h_{\theta }\le 1$$

g is the logistic or sigmoid function, given as below.

(20) $$g(z)={\displaystyle \frac{1}{1+e^{-}z},wherez={\theta }^x}$$

A loss function is defined to know the performance of the algorithm using the weights where,

(21) $$h=g(X\theta )J(\theta )={\displaystyle \frac{1}{m}.(-y^T\log (h)-(1-y{)}^T\log (1-h))}$$

After this, loss function is minimized by decreasing and increasing the weights. The gradient descent tells us how the loss changes with the modification of parameters.

(22) $${\displaystyle \frac{\delta J(\theta )}{\delta {\theta }_j}={\displaystyle \frac{1}{m}{X}^T(g(X\theta )-y)2}}$$

Multinomial logistic regression Here 3 or more possible un-ordered types of dependent variables are present with no quantitative significance, such as different types A, B or C. The MLR model is an extension of LR (logistic regression) model. LR Model:

(23) $$\pi (x)={\displaystyle \frac{\alpha +{\beta }_1{x}_1+{\beta }_2{x}_2+\cdots +{\beta }_p{x}_p}{1+e^{\alpha +{\beta }_1{x}_1+{\beta }_2{x}_2+\cdots +{\beta }_p{x}_p}}}$$π(x) is an event probability, α represents the dependent variable, β₁, β₂ are independent variables, x₁, x₂… are regression coefficients, p is the number of independent variables and e is the term representing the error.

From this LR model, MLR can be obtained as an extended version as below. MLR Model:

(24) $${\pi }_j(x_i)={\displaystyle \frac{e^{{\alpha }_i+{\beta }_{1}j{x}_{1}i+{\beta }_{2}j{x}_{2}i+\cdots +{\beta }_{p}j{x}_p}}{1+\sum _{j=1}^{k-1}{e}^{{\alpha }_i+{\beta }_{1}j{x}_{1}i+\beta 2jx2i+\cdots +{\beta }_{p}j{x}_p}}}$$

here j₁, j₂..j_k are the k categories and n (i₁, i₂…i_n) are the possible independent levels.

Ordinal regression

It deals with a quantitative significance having 3 or more ordered types for the dependent variable. As an example, variables can be “poor”, “good”, “very good” or “excellent” categories, can be represented as a score of 0 to 3.

Naïve Bayes algorithm

It is based on Bayes’ theorem and is a probabilistic algorithm. It assumes the occurrence of a feature independent of the other features. Bayes theorem forms its basis,

(25) $$P(Y|X)={\displaystyle \frac{P(X|Y)P(Y)}{P(X)}}$$

It gives the relation between an arbitrary event X with y causing due to some random process. Here P(X) is the input probability, and P(Y) is considered as output probability, while P(Y--X) defines the probability of Y states versus the X input probability.

The major disadvantage of this algorithm is, it assumes all features as independent features. Hence not possible to get the relationship between the features.

• It has three types viz Gaussian, Multinomial, and Bernoulli.

• Gaussian Naive Bayes In this type, features have values that are continuous in nature and are considered to have Gaussian distribution.

• Multinomial Naive Bayes: In this case, events are considered to occur in multinoial distribution type. The feature vector is formed by the frequencies of occurrence of some events. Typical use includes the document classification.

• Bernoulli Naive Bayes: Here, input describes the features with a binary variable independent of each other. In this model, features used are binary term occurrence instead of frequencies.

Support Vector Machine (SVM)

Support vector machine (SVM) is a machine learning algorithm that has a supervised type. It finds its application in classification as well as regression. It has two types. Linear SVM, used for linearly separable data and non-linear SVM, is used for the data that is non-linearly separable; a non-linear type of SVM is used. Figure 7 explains the SVM and its different concepts. The followings are important terms in SVMs.

Hyperplane is the space or plane for decision. It has a division in a set of different objects having separate classes.

Support Vectors These are the data points which are closest to the hyperplane. These data points define the separating line.

Margin The gap between two lines on Closest data points is considered to define the margin. It is given by the gap in the lines of different classes. The perpendicular distance between the line and the support vectors is used to find the margin. Larger margins are good, and smaller ones prove to be bad.

Convolutional neural networks

Figure 8 shows the generalized architecture of convolutional neural network (Phung & Rhee, 2019). The CNN has two basic building blocks:

The Convolution Block—Consists of two layers, namely, the Convolution and the Pooling Layer. For feature extraction, these layers prove to be the essential components.

The Fully Connected Block—has a simple neural network that is fully connected architecture. This layer’s function is the classification based on the input coming from the convolutional block.

Pooling Layer—It has the process where the extraction of a value is done from a set of values. It generally uses the maximum or the average value. This reduces the output matrix size. There are some popular architectures of CNN given as, Classic network architectures viz LeNet-5, AlexNet, VGG 16, and Modern architectures namely Inception, ResNeXt, ResNet, DenseNet.

BCCD Database

Total images in this data are 12,500. Each subtype, Eosinophil, Lymphocyte, Monocyte, and Neutrophil, has approximately 3,000 images. This Dataset also includes additional 410 images of WBC and RBC with JPEG and xml metadata format. This database is under MIT license (Yildirim & Çinar, 2019).

ALL-IDB (Acute Lymphoblastic Leukemia Database) This database is for the analysis and research on the disease of acute leukemia. This dataset has two subtypes ALL-IDB1 and ALL-IDB2. All images are taken with Canon Powershot G5camera. The microscope has 300 to 500 magnification. Images follow the JPEG format with a color depth of 24-bit. ALL-IDB1 dataset has 109 images with a resolution of 2,592 × 1,944 and a total element of 3,900. It has a total of 510 lymphoblasts. ALL-IDB2 dataset has 260 images with a resolution of 257 × 257 and a total element of 257. It has a total lymphoblast of 130 (Labati, Piuri & Scotti, 2011; Donida Labati, Piuri & Scotti, 2011).

Atlas of hematology by Nivaldo Mediros

This database contains different types of microscopic blood cell images. Different image types include maturation sequence, alteration of erythrocytes, anemia images, leukemia images, parasites, fungus images, and some miscellaneous images. The photos of images were taken with the magnification of 200, 400, 360, and 1,000 by the photo-microscopes of Zeiss and Nicon (Acharya & Kumar, 2019).

ASH image bank

This database is from the American Society of Hematology (ASH). It is basically an online image bank of having leukemia cells images. This image library web-based. This offers a wide collection of hematology categories. They provide images with variations in resolutions (Agaian, Madhukar & Chronopoulos, 2014).

Leukocyte Images for Segmentation and Classification (LISC)

This dataset contains images of healthy subjects. The total number of images in the dataset is 400 from 100 different blood slides of healthy eight subjects. The size of the images is 720 × 576. These images are from Hematology-Oncology and BMT Research Center of Imam Khomeini Hospital in Tehran, Iran. A total of 250 images are available for experimentation purposes. Images have different elements of leukocyte like eosinophil, basophil, lymphocyte, monocyte, and neutrophil (Rezatofighi & Soltanian-Zadeh, 2011; Livieris et al., 2018).

C-NMC dataset

This database is owned by The Cancer Imaging Archive (TCIA) public access. This database has a total of 118 participants, with a total of 15,135 images. All images are in bit map image (BMP) format. This data is divided into train sets, preliminary test sets, and final test sets with different cancerous and normal images. Table 1 shows the comparison of different publicly available datasets, and Table 2 gives the information of datasets availability along with urls.

Table 3 explores the different methods used by researchers along-with the performance measures, datasets, and a number of images used for disease detection.

Discussion of explainable AI for critical analysis

In medical imaging, artificial intelligence is mostly used for detection and diagnosis (Ahuja, 2019; Tizhoosh & Pantanowitz, 2018). It is generally not been preferred as the final decision about the diagnosis. The main reason for this is, it works as a black box with only input and outputs are known. There might be the occurrence of the right decisions related to the diagnosis, but it might be due to wrong reasons. The algorithms such as decision trees elaborate their implementation to a good extent. However, it limits the parameters such as accuracy of diagnosis, although advanced AI algorithms and deep learning algorithms assure good accuracy in terms of diagnosis but unable to explain the insides of implementation-black box (Lundervold & Lundervold, 2019). Hence, explainable AI came into the picture to justify the trust of diagnosis in medical imaging (Singh, Sengupta & Lakshminarayanan, 2020). This will analyze the black box’s features and characteristics in the implementation of AI or deep learning algorithms. In order to make the system to be used with more trust for commercial purposes for the general public, the explainability will prove to be the most suitable and appropriate.