Multimed Tools Appl DOI 10.1007/s11042-017-4478-3

Classification of acute lymphoblastic leukaemia using hybrid hierarchical classifiers Jyoti Rawat 1 & Annapurna Singh 1 & H. S. Bhadauria 1 & Jitendra Virmani 2 & J. S. Devgun 3

Received: 30 August 2016 / Revised: 20 December 2016 / Accepted: 3 February 2017 # Springer Science+Business Media New York 2017

Abstract In current consequence of haematology, blood cancer i.e. acute lymphoblastic leukemia is very frequently founded in medical practice, which is characterized by over activation and functional abnormality of bone marrow. The abnormality is identified through physical examination with a screening of blood smears. However, this method is error prone and labor intensive task for haematologist. Hence, haematologist needs a specific computer aided diagnostic system (CAD) that can deal with these limitations of prior systems and capable of discriminating immature leukemic cells from mature healthy cells. Thus, this work addresses the problem of segmenting a microscopic blood image into different regions, and then further analyzes those regions for localization of the immature lymphoblast cell. Further, it investigates the use of different geometrical, chromatic and statistical textures features for nucleus as well as cytoplasm and pattern recognition techniques for sub typing immature acute lymphoblasts as per FAB (French– American – British) classification. This can facilitate haematologist for acquiring essential information about prognosis and for an appropriate cure for leukemia. The exhaustive experiments have been conducted on 260 microscopic blood images (i.e. 130 normal and 130 cancerous cells) taken from ALL-IDB database. The proposed techniques consisting of the segmentation module used for segmenting the nucleus and cytoplasm of each leukocyte cell, feature extraction module, feature dimensionality reduction module that uses principal component analysis (PCA) to mapped the higher feature space to lower feature space and classification module that employs the standard classifiers, like support vector machines, smooth support vector machines, k-nearest neighbour, probabilistic neural network and adaptive neuro fuzzy inference system.

* Jyoti Rawat [email protected]

1

G.B. Pant Engineering College, Garhwal, Pauri, UK 246001, India

2

CSIR-Central Scientific Instruments Organization, Chandigarh, India

3

M. M. Institute of Medical Sciences & Research, Solan, HP, India

Multimed Tools Appl

Keywords Leukaemia . Statistical texture features . Chromatic features . Geometrical features . Principal Component Analysis . Hierarchical classifiers

1 Introduction Leukaemia is a kind of blood malignancy that is characterized by unrestrained accretion of immature and non-functional leukocyte cells in blood-forming organs like lymph tissue or stem cells of bone marrow that is a centre of most bones where blood cells are formed. The main risk factors for blood cancer are exposure to electric magnetic fields, genetic conditions, extraordinary doses of radiation or exposure to some chemicals solvents [4]. The relative survival rates against leukaemia i.e., blood cancer vary according to a person’s age, gender, race and subtype of leukaemia [2]. The statistics reported in different articles during 2004 to 2010 shows that the overall relative survival rate was 60.3%. From 2004 to 2010, the five-year relative survival rates overall were 59.9% for CML, 83.5% for CLL, AML 25.4% for overall and 66.3% for adolescents than 15 years, for ALL 70% overall, 91.8% for children and teenagers younger than 15 years, and 93% for children younger than 5 years. According to fact spring 2015, leukaemia and lymphoma society [2] an estimated 52,380 new cases of leukaemia are expected to be diagnosed in the US in 2014. Cases of acute leukaemia are expected to account for 14.7% more cases than chronic leukaemia. Basically, blood disorders can be classified on the basis of their cause and cell of origin as hereditary, infectious, endocrine (hormone glands), neoplastic (tumour), based on the immune system, and traumatic [33], etc. as shown in Fig. 1. Among all malignant neoplastic disorder, leukaemia (blood cancer) is in the research forefront for several medical researchers. As shown in Fig. 1, leukaemia’s and lymphomas both are a group of tumours that affect the hematopoietic and lymphoid tissues of blood, bone marrow, and lymphoid system. Leukaemia can be divided into four main types, chronic myeloid leukaemia (CML), acute lymphoblastic leukaemia (ALL), chronic lymphocytic leukaemia (CLL) and acute myeloid leukaemia (AML). The terms Bmyeloid^ or Bmyelogenous^ and Blymphoid,^ or Blymphoblastic^ denote the cell types involved. ALL is characterized by a rapid enhancement in the number of immature blood cells. Standard promising treatments involve chemotherapy and radiotherapy [27]. The WHO and French – American –British (FAB) are two common standards used widely for acute lymphoblastic leukemia classification. According to FAB classification, there are subtypes of ALL: L1, L2 and L3 which are based on the morphological structure of cell [4]. As per the visual inspection and information provided by hematopathogist, it is found that shape-size of these types of acute leukemic cells are irregular that cannot be easily visualized by the microscopic examination. Hence, it is a mind-numbing task for the haematologist to early prediction of ALL because of its rapid expansion in stem cell and can be fatal if untreated. Therefore, an expert system [59–64] is required for the complementary blood examination that must be more precise, inexpensive, efficient and expeditious. The analysis of cancerous ALL lymphoblast’s is quite tedious because of its variant nature for its shape, colour, dimensions and texture as shown in Fig. 2.

Multimed Tools Appl

Fig. 1 Analytical description of haematological disorders of leukocytes. Note: Highlighted portion shows that the work is done in present paper

2 Related works As per the study of literature, it has been found that the existing systems are only partially automated [23, 24, 28, 30, 44, 46, 49, 53, 58] and are moderately able to analyze the leukocyte and detect malignancy from microscopic images. The detailed explanation of literature and work carried for classification of normal and cancerous microscopic leukocyte cells is tabulated in Table 1.

Multimed Tools Appl

Fig. 2 A comparison between different type of lymphoblast suffering from leukaemia according to FAB and WHO classification as L1, L2, L3 and Precursor-B, Precursor-T [4]

It can be observed that most of the work [1, 16, 31, 34, 35, 41–43, 45, 56] for classification of leukocytes for leukemic cell reorganization is emphasized on various texture models, some studies [6, 16, 31, 34, 41–43, 56] have been carried out on colour based features and some focused on shape models [1, 6, 29, 31, 41, 45] for classification of peripheral blood smear images. From the related work it is a specific information observed that different type of leukocytes are differed in their morphological appearance in blood smear, and infected white cell that is known as lymphoblast cells or leukemic cells can also be discriminated from the healthy white cells (lymphocytes) by analyzing shapes, morphology and colour of their cytoplasm and nucleus region present in leukemic cell [38, 52, 65]. It can be examined from the Table 1, that there are various studies has been done for the detection of cancerous cell. The maximum classification accuracy for 2-class classification problem it is 98.2%. In this work [42], artificial neural network classifier is used; various cytological and laboratorial features are extracted. The detailed explanation of work carried for the FAB classification (L1, L2, and L3) of acute lymphoblastic leukemia (ALL) leukocyte cells is tabulated in Table 2. Similarly, in Table 2, it can be seen that 98.0% is the highest accuracy achieved for 3-class ALL (FAB) classification by using PCA-SVM classifier by extracting shape and texture features [39]. From the extensive study of literature, it can be seen that there are few studies related to FAB class classification of ALL with the public data set, so there is a need for a consistent, reliable Computer Aided Diagnosis (CAD) system that must belong to a publically available dataset where others can validate their systems. Hence present work proposes a hybrid hierarchical Computer Aided Diagnosis (CAD) system that analyzes microscopic blood images for the early prediction of ALL that will discriminate the cancerous cell (CC) from the normal cells (NC) and simultaneously sub-type the ALL. It is worth mentioning that the hierarchical CAD designs provide the possibility to go stepwise from the general classification problem, i.e., CC versus NC. Further, CC cells are classified into their subtypes as per FAB classifications with the hierarchical framework of classifiers. The main objectives of this work are (i) finding true edges of nucleus and cytoplasm, (ii) correct classification of cancerous malignant lymphoblast cells and normal cells, and (iii) correct classification of FAB classes of ALL (L1, L2, and L3). To achieve these objectives, components of leukocyte cell which are nucleus and cytoplasm are segmented from the cell background. Then various geometrical, chromatic, and statistical texture features are extracted from segmented cytoplasm and nucleus for an accurate classification of acute lymphoblast cell. After the feature extraction step, different classifiers (PCA-kNN, PCA-PNN, PCA-SSVM, PCA-ANFIS and PCA-SVM) in hierarchical order are used to recognized lymphocytes as cancerous i.e., lymphoblast or not and discriminate subtypes of ALL.

Shape and texture features Texture features Shape features Shape, colour, and texture features Shape and texture features Shape, texture, and colour features Shape features Shape and texture features Shape and texture features Shape and texture features Shape, colour, and texture features Shape and colour features Shape and texture features Multiple clinical, laboratorial features Shape and texture features Shape and texture features Shape, texture, and colour features Shape and texture features Fractal dimensions, colour, and shape features Shape features

Singh, G. et al. 2016, [55] Singhal, V. et al. 2016, [57] Zhang, L. et al. 2016, [68] Viswanathan, P. et al. 2015, [65] Amin, M.M., et al. 2015, [3] Neoh, S.C., et al. 2015, [41] Bhattacharjee, R. et al. 2015, [6] Singhal, V. et al. 2015, [56] Rawat, J. et al. 2015, [48] ElDahshan, K. et al. 2015, [13] Mohapatra, S. et al. 2014, [36] Lorenzo Putzu et al. 2014, [45] Aimi, A.N. et al. 2013, [1] Putzu, L. et al. 2013, [43] Pedreira, C.E et al. 2012, [42] Madhloom, H.T et al. 2012, [29] Mohapatra, S. et al. 2012, [35] Mohapatra, S.et al. 2011, [34] Halim, N.H.A. et al. 2011, [16] Mohapatra, S. et al. 2010, [31] Mohapatra, S. et al. 2010, [32] Scotti, F. et al. 2005, [52]

ANN SVM Fuzzy system Fuzzy system SVM Dempster-Shafer ANN SVM SVM Field Programmable Gate Arrays EOC* SVM MLP_BR SVM ANN kNN ANN SVM Cell counting based SVM SVM FFNN

Classifier used

EOC- Ensemble of Naive Baiyes, kNN, MLP-SVM, RBFN-SVM and SVM classifiers

Feature extraction method

Authors, Year

Table 1 Studies carried out for the classifications of normal and cancerous leukocyte cells

ALL-IDB ALL-IDB Harbin Medical University, China ALL-IDB Omid hospital ALL-IDB ALL-IDB ALL-IDB ALL-IDB ALL-IDB Ispat General Hospital, Rourkela, India ALL-IDB ALL-IDB Federal University of Rio de Janeiro (UFRJ) University of Malaya, Malaysia Ispat General Hospital, Rourkela, India Ispat General Hospital, Rourkela, India Hospital Universiti Sains Malaysia, Kelantan Ispat General Hospital, Rourkela, India Ispat General Hospital, Rourkela, India M. T Research Centre, Italy

Dataset used 108 260 108 21 180 120 260 196 300 104 267 230 245 189 260 100 108 50 108 108 113

Image

97.2 93.8 98.0 97.0 96.7 95.2 92.3 89.8 94.7 92.0 95.7 92.0 98.2 92.5 93.0 97.8 95.0 95.0 75.9

Accuracy (%)

Multimed Tools Appl

Multimed Tools Appl Table 2 Studies carried out for the FAB classification (L1, L2, and L3) of acute lymphoblastic leukemia (ALL) leukocyte cells Authors, Year

Feature extraction method Classifier used

MoradiAmin, M.et al. 2016, [39] Shape and texture features MoradiAmin, M. et al. 2015, [38] Shape and texture features Mohapatra, S.et al. 2014, [37] Color, shape, and texture features

Dataset used

Image Accuracy (%)

PCA-SVM Omid hospital 21 SVM Omid hospital 312 Ispat General Hospital, 67 EOC* Rourkela, India

98.0 97.0 97.3

EOC- Ensemble of Naive Baiyes, kNN, MLP-SVM, RBFN-SVM and SVM classifiers. L1, L2, and L3 are the FAB classes of Acute lymphoblastic leukemia (cancerous cell)

3 Material and methods 3.1 Database description The dataset in the present work contains 260 ROI images taken from ALL-IDB 2 provided by M. Tettamanti research centre, Monza, Italy [25]. The ALL-IDB database has two discrete versions are ALL-IDB1 and ALL-IDB2 labelled by expert oncologists. In the present work, the balanced and imbalanced dataset is considered. This balanced dataset contains 130 normal cells (NC) and 130 cancerous cells (CC). Among 130 cancerous samples, first half used for train and next half is used for testing the classifier. Similarly, in the case of healthy lymphocyte images, the first half is used for training purpose and next half is used to validate the classifier. Further 130 cancerous cells are labelled as per FAB classification. By the participating pathologist who has more than 20 years of experience in haematology department classified the ground truth into three class’s i.e. L1, L2, and L3. The L1, L2, and L3 classes contain

Fig. 3 Test microscopic blood images of a normal cell (NC) and a cancerous cell (CC) randomly taken from the ALL-IDB 2. Note: NC: Normal cells, CC: cancerous cell and L1, L2, L3 are FAB classes

Multimed Tools Appl

79, 39, and 10 cells respectively. The remaining two cells are considered as the atypical case. Database description for malignant lymphoblast classification and brief information regarding dataset and bifurcation of training and testing of L1, L2, and L3 are given in Table 14. The sample images of a normal cell (NC) and a cancerous cell (CC) randomly taken from the ALL-IDB 2 dataset are shown in Fig. 3.

3.2 Proposed CAD system The flow chart for analysis of microscopic blood images for early detection of acute lymphoblastic leukemia is shown in Fig. 4. It comprises of segmentation module, feature extraction module, and classification module. In this CAD system, segmentation module is used for the extraction and identification of nucleus and cytoplasm from lymphocyte cell. To find out the morphological differences between leukemic and healthy lymphocyte cell, a different characteristic of their chromatic, geometric and texture are extracted by feature extraction. These extracted features are processed by the machine learning module known as a classifier to detect the cancerous and normal lymphocyte. The brief explanation of each module is given in next section of this document.

Fig. 4 Proposed Hybrid hierarchical classification system

Multimed Tools Appl

3.2.1 Segmentation module Segmentation of blood microscopic images is the basis for all automated blood malignancy identification and classification methods based on image analysis. In the present work, intensity based segmentation is used; pixel gray level value is the basic source of information for an entity. Therefore, this step segments the objects of interest (nucleus and cytoplasm) from the microscopic image of leukocyte using different morphological operations [14, 47, 54]. The segmentation module as shown in Fig. 5, produces the segmented nucleus and cytoplasm of the leukocytes, which can be processed by the feature extraction module for the feature extraction and extracted feature vector set are used for the characterisation of the cancerous cells, L1 cells, L2 cells and L3 cells. The brief steps for segmentation of lymphocyte cell are described in Algorithm 1. Algorithm 1 Step 1: First step is to acquire an image from a benchmark dataset ALL-IDB2 [45] which contains the cropped ROIs of WBC. Step 2: Transform the colour input (RGB scale) image into gray scale image. Step 3: Then apply pre-processing to enhance the image quality using histogram equalization and to reduce the noise by using 2-D order statistical filter {Replace each pixel of the input image by non-zero pixel in the sorted set of neighbours}. Step 4: Apply the global thresholding on pre-processed image of lymphocyte for extracting different objects (nucleus and cytoplasm) with different pixel intensity. Step 5: The Morphological opening is applied on threshold image that will even the contour of the lymphocyte cell and eliminates the pixels that cause objects overlapping. Step 6: After the morphological reconstruction of the image, segmented nucleus is obtained. Step 7: Subtract the segmented nucleus from the pre-processed image to obtain the cytoplasm.

Basically, it’s a two parts process of leukocyte segmentation. First is segmentation of the nucleus component and second is cytoplasm segmentation after converting input cell image into gray scale image. The original input image usually effected with a different type of noise background and outliers (inappropriate staining, illumination variations or variability in the ROI) that make segmentation vital and difficult process. To overcome these noises image need to be pre-processed before analyzing so some pre-processing techniques for image enhancement and order statistic filter are used. The brief description and results are discussed below:

Pre-processing The pre-processing module includes histogram equalization to enhance the quality of image and order statistic filter which is used to reduce the noise as well smoothing the image. Each step is described below and the result of each stage is shown in Fig. 6. Histogram equalization Let pr(r) denote the probability density function (PDF) of the normalized intensity levels in a given image. The transformation on the input mage to obtain desired resultant image s, r

s ¼ T ðrÞ ¼ ∫0 pr ðωÞdω

ð1Þ

here ω is a dummy variable. The probability density function of the output image is ps ðsÞ ¼


1 for0 ≤ s ≤ 1 0 otherwise

ð2Þ

In general, the histogram of the filtered image will not be uniform, due to distinct nature of variables. Those filtered images have the low dynamic range and improper contrast, the preceding transformation that is cumulative distribution function produce an image whose

Multimed Tools Appl

Fig. 5 Overview of segmentation of cytoplasm and nucleus

intensity levels are uniform likely. The result of this, intensity level equalization process is an image with high dynamic range, which will tend to have higher contrast [14].

Order statistic filter Microscopic blood images are acquired by a digital camera that captures the images in a different environment and can be operator dependent. The images may have non-uniform illumination because of variation in colour and different contrast among the different region of the same image. Therefore order statistic filter is used to normalize the image. Order statistic filter [12] also known as the rank filter is nonlinear spatial filter which replaces the value of the centre pixel with the value obtained by ranking in the neighbourhood. If the order is one than the statistic filter is known as min filter and the second order is known as a median filter. The filtering operation can be further characterized through the determination of output distributions of the input image. Assume that input image consists of noise sample with probability distribution fx(⋅) and cumulative distribution (cdf) Fz(⋅). Under this situation, the median filter outcome cdf as Fmed(⋅), and pdf, fmed(⋅), are given by F med ðtÞ ¼

N1

∑

i¼N 1 þ1

N  i

F x ðtÞi ð1−F x ðt ÞÞN −i

ð3Þ

Multimed Tools Appl

Fig. 6 The result of Preprocessing of Leukocyte microscopic blood images: (a-e) original images from a dataset, (f-j) rgb to gray conversion of input images, (k-o) enhanced images after histogram equalization,(p-t) filtered images with uniform illumination

and f med ðtÞ ¼

N! f ðtÞ F x ðtÞN 1 ð1−F x ðt ÞÞN 1 N 1 !N 1 ! x

ð4Þ

Nucleus segmentation Leukocyte nucleus has inconsistent shapes based on different kind of leukocytes. So it is hard to find a significant segmentation method to characterize its shapes and size. The cell segmentation methods can be categorized into threshold-based, patter recognition-based methods, contour-based method, and metaheuristic-based method (genetic algorithm, differential algorithm, an artificial bee colony and micro canonical annealing) [51]. Among all the segmentation methods, threshold based methods are fast and reliable for uniform images where the nucleus and cytoplasm have almost homogeneous areas with respect to the gray level intensity. Global Thresholding. The global thresholding [14] is used to detect the precise edge of the nucleus in leukocyte cell. Minimization of intra-class variance is used in the global thresholding method to select the threshold. The intra-class variance is computed as: σ2w ðt Þ ¼ ω1 ðt Þσ21 ðt Þ þ ω2 ðtÞσ22 ðtÞ

ð5Þ

Multimed Tools Appl

Fig. 7 Result of the nucleus and cytoplasm detection: (a-e) Images after global thresholding, (f-j) Images after morphological opening, (k-o) nucleus extraction, (p-t) cytoplasm extraction

here threshold t separates the probability of the two classes represented as ω1 and ω2 with their variances termed as σ21 and σ22 , respectively. After thresholding, the morphological opening will be done on the threshold output. Morphological Opening. For morphological reconstruction, morphological opening [14] is applied by using structuring element (SE) which is a shape that has a particular size. Morphological opening obtained by first applying the erosion function E(x,y) on input image i(x,y) by SE(x,y).
  E ðx; yÞ ¼ iðx; yÞΘSEðx; yÞ ¼ z SEðx; yÞz ⊆fiðx; yÞg

ð6Þ

here i(x,y) is binary image and SE(x,y) is a structuring element in the X-Y plane. It means corrodes the image pixels that are smaller than the SE(x,y). Then obtained results is dilating D(x,y) by SE(x,y) that expand the object in the image based on the shape and size of SE(x,y). The combination of followed by dilation known as morphological opening described in eq. 8.
   Dðx; yÞ ¼ iðx; yÞ⊕SEðx; yÞ ¼ z SEðx; yÞz ∩iðx; yÞ ⊆iðx; yÞ

ð7Þ

MO ¼ E ðx; yÞ∘Dðx; yÞ ¼ ðiðx; yÞΘE ðx; yÞÞ⊕Dðx; yÞ

ð8Þ

Multimed Tools Appl

here MO is morphological opening that will even the contour of the lymphocyte cell and final segmented nucleus is obtained.

Cytoplasm segmentation Cytoplasm area is achieved by deducting the segmented nucleus from the input leukocyte image. The final nucleus and cytoplasm portion obtained from segmentation process are shown in the Fig. 7.

3.3 Feature extraction module The prominent extraction of features of the objects plays an essential part of increasing the performance and reducing the complexity of a classifier [40]. Generally, the recent literature mostly focused on morphological [67], chromatic [17] and texture [5, 7, 18, 50] features extracted from particular blood cell [6, 13, 36, 41, 45, 48, 56]. The brief description of the morphological difference between cancerous white cells that is stated as lymphoblast and normal white cells (leukocytes) are given in Table 3. The same fact has been also observed by the participating radiologist during the diagnosis of blood disordered and also noticed that the cytoplasm, as well as nucleus, plays a very crucial role for detecting ALL. Accordingly, in this study, eleven geometrical features, fifteen chromatic features and forty-five statistical texture feature using first order, second order, and higher ordered models are computed from cytoplasm and nucleus of each cell. The brief descriptions of computed features are given in Table 4.

3.4 Feature space dimensionality reduction using PCA (principle component analysis) The present classification problem employed the PCA for reduction of large features space dimensionality that utilizes the structure of the principal components of a feature set to find a subset of the original feature vector. PCA derive a linear projection of high-dimensional data into a lower dimensional subspace such as the variance retained is maximized and the least square reconstruction error is minimized [5, 10, 19, 21]. To reduce dimensionality from d to m PCA performs following steps: i) First, centre the data and take off the mean. ii) The d × d covariance matrix is calculated.

C¼

AT A N

ð9Þ

iii) The eigenvectors of the covariance matrix (orthogonal) are calculated. iv) Select the m eigenvectors that match up to the heights m eigen-values to be the new space dimensions. The variance in each new dimension is given by the eigen-values, a eigenvalues problem like:

Oval Indented

Blue-purple Red-purple

Less Large

Light sky blue Dark blue

Missing Clogged Discrete Untie

Soft Coarse

Rough Velvety

N/C ratio Nucleus shape Nucleus colour Nucleus size Cytoplasmic colour Nucleoli Nuclear chromatin Nucleus boundary Cytoplasmic boundary

Normal lymphocytes cell (NC) Stumpy Cancerous cell (CC) Elevated

Attributes (Features)

Table 3 Morphological discrimination between cancerous lymphoblast (CC) and normal lymphocytes (NC) cell

Multimed Tools Appl

Multimed Tools Appl Table 4 Various texture and geometrical descriptors extracted for early cancer detection Feature Method

Extracted features

lc

ln

Geometrical Features

Area, Perimeter, Diameter, Euler’s no, Major axis, Minor axis, Solidity, Eccentricity, Roundness, Convex area and Extent. MV_red, MV_green, MV_blue, STDV_red, STDV_green, STDV_blue, MOMENTS3_red (skewness), MOMENTS3_green, MOMENTS3_blue, MOMENTS4_red (kurtosis), MOMENTS4_blue, MOMENTS4_green, MOMENTS5 (high order moments)_red, MOMENTS5_green and MOMENTS5_blue. Smoothness, Standard deviation, Third moment, Uniformity, Average gray level, FOS, and Entropy. Variance, Contrast, Correlation, Entropy GLCM, Inverse difference moment, Sum average, Sum variance, Sum entropy, Difference entropy, Difference variance, Angular second moment, Information measures of correlation-1 and Information measures of correlation-2. Mean and Variance. Contrast, Coarseness, Periodicity and Roughness. Entropy, Homogeneity, Energy, Contrast and Mean. Contrast, Coarseness, Business, Complexity and Strength. Low gray level run emphasis, High gray level run emphasis, Run percentage, Short run low gray level emphasis, Short run high gray level emphasis, Long run low gray level emphasis, Long run emphasis, Long run high gray level emphasis, Gray level non-uniformity, Short run emphasis, Run length non-uniformity and.

11

11

15

15

6

6

13

13

2 4 4 5 11

2 4 4 5 11

Chromatic features

Statistical Texture models

FOS GLCM

EDGE SFM GLDSM NGTDM GLRLM

Total

142

MV Mean value, STDV Standard deviation value, FOS First order statistics, GLCM Gray level co–occurrence matrices, SFM Stastical feature measure, GLDSM Gray level difference statistics, NGTDM Neighbourhood graytone difference matrix, GLRLM Gray level-run length matrix, lc length of feature vector for cytoplasm, ln length of feature vector for nucleus

&

Find the direction for which the variance is maximized: r1 ¼ argmaxr1 varðYr1 Þ T

subject : r1 r1 ¼ 1

&

ð10Þ

Rewrite in terms of the covariance matrix: var ¼ N

−1



Y ̂r1

T

! Y ̂Y ̂ r1 ¼ r1 T Sr1 T

Y ̂r1 ¼ r1

T

N

−1

! Y ̂Y ̂ ¼ sc

ð11Þ

T

N

−1

ð12Þ

here, sc is sample covariance.

&

&

&

Solve via constrained optimization:



T Lðr1 ; λ1 Þ ¼ r1 T Sr1 þ λ1 1− r1 r1

ð13Þ

Gradient with respect to r1

Multiply by

dLðr1 ; λ1 Þ ¼ 2Sr1 −2λ1 r1 ⇒Sr1 ¼ λ1 r1 dr1

ð14Þ

r1 T ¼ λ1 T Sr1

ð15Þ

Multimed Tools Appl

The projection variance of each principal component is given by its eigen-value. For this classification problem, applied the feature selection for every possible subset of the feature set of different length of the feature vector for statistical texture-based feature as well as geometric features. To obtain the optimum number of principal components, extensive experiments have been carried out for each classifier by varying the number of pc values from 2 to 15 i.e. firstly taken two principal component then first three principal components and so on.

3.5 Classification module Classification process includes a large series of decision-theoretic approaches to identify the image of an interest. The entire features extracted from above phase are used here for classification purpose. Each of these extracted features belongs to one of prior specified distinct classes in case of supervised classification or clustered into sets of archetype classes in case of unsupervised classification [11, 15, 66]. The classification module of the present work is consisting of two modules, (a) Classification module one: This module is used for the detection of cancerous leukemic cells (CC) from the normal lymphocyte cell (NC), in which cell having a complex background, a vast number of outliers or noise due to staining can be generated with the prominent features. They increase the computational complexity of classifier and reduce the performance. (b) Classification module two: This module is used for the classification of ALL or cancerous cells (CC) into L1, L2, and L3 as per FAB classification. The brief description of the hierarchical framework of classification module is given in Fig. 8. To implement the hierarchical classification module for classification of normal cell versus cancerous cell, five classifiers (i.e. PCA-kNN, PCA-PNN, PCA-SVM, PCASSVM, and PCA-ANFIS) are arranged in hierarchical order at three levels. The

Fig. 8 Architecture of general hierarchical framework of classification module

Multimed Tools Appl

classifier 1 at level 1 is used for the classification of normal cells versus cancerous cells. If the prediction of test input cell is cancerous then it is inputted to classification module 2. Classification module 2, which is consisting of classifier 2 and classifier 3 is used for the FAB classification. Classifier 2 at level 2 is used for the classification between L1 versus remaining class. If the prediction of the input cell is remaining class then it is inputted to classifier 3 at level 3, which is used for the classification between L2 versus L3.

3.5.1 Design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA- kNN classifiers. The k-NN classifier [29] is applicable where there is a less knowledge about the distribution of the dataset. It is based on the Euclidean distance between a test sample and the training sample of the dataset. 

d ai ; a j



qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2  2  2 ¼ ai1 −a j1 þ ai2 −a j2 þ ::::::: þ ain −ajn :

ð16Þ

here d(ai,aj) is an Euclidean distance between ai [(ai1), (ai2), (ai3), (ai4),.... (ain)] and aj [(aj1), (aj2), (aj3), (aj4),.... (ajn)] feature set. The architecture of hierarchical classification of NC/CC cell and FAB classification of acute lymphoblastic leukaemia using PCA- kNN classifier is arranged in the same manner as Fig. 8.

3.5.2 Design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA- PNN classifiers. The PNN [15] is predominantly a classifier which performs a mapping from input pattern to different classifications. It is an execution of kernel discriminant analysis statistical algorithm which structured into the multilayer feed forward network. It follows Bayes optimal decision rule. hi ci f i ðX Þ > h j c j f j ðX Þ∀ j≠i

ð17Þ

here, hi is the previous probability of unknown samples and ci is misclassification cost. The architecture of hierarchical classification of NC/CC cell and FAB classification of acute lymphoblastic leukaemia using PCA-PNN classifier is arranged in the same manner as Fig. 8.

3.5.3 Design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA-SSVM classifiers. The Smoothing classification method is widely used for solving essential primal programming problems and applications. SSVM classifier is implemented using the SSVM toolbox developed at Data Science and Machine Intelligence laboratory, Taiwan [26]. The architecture of

Multimed Tools Appl

hierarchical classification of NC/CC cell and FAB classification of acute lymphoblastic leukaemia using PCA-SSVM classifier is arranged in the same manner as Fig. 8.

3.5.4 Design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA- ANFIS classifiers. An ANFIS classifier [22] is a multilayer feed-forward network consisting of the input layer, membership layer, fuzzification layer, defuzzification layer, normalization layer, and output layer [22]. The architecture of hierarchical classification of NC/CC cell and FAB classification of acute lymphoblastic leukaemia using PCA-ANFIS classifier is arranged in the same manner as Fig. 8.

3.5.5 Design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA-SVM classifiers. It has been reported in [10, 18, 23] that machine learning system which uses classifier like support vector machines are best suited and obtain good characterization to a certain degree. The main task of the SVM classifier [8, 9, 20, 26] is to finding the optimum hyper plane in the higher dimensional feature space to separate the training data with one class to other class. For this binary classification problem, LibSVM library is used a binary classifier SVM (Normal cell/cancerous cell). The architecture of hierarchical classification of NC/CC cell and FAB classification of acute lymphoblastic leukaemia using PCA-SVM classifier is arranged in the same manner as Fig. 8.

3.6 Classification performance measurement For the assessment of designed classifiers classification performance, the system is evaluated by medically significant statistical measures as overall classification accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). These statistical measures are briefly explained below (Table 5). Here TP (true positive), TN (true negative), FP (false positive) and FN (false negative) can be defined as: TP: TN: FP: FN:

These are healthy identified cells. These are correctly identified the cancerous cell. These are misclassified normal objects. These are misclassified cancerous objects.

Table 5 Description of performance Classification result

Normal Cancerous

Haematologists opinion Normal

Cancerous

TP FN

FP TN

Multimed Tools Appl

3.6.1 Overall classification accuracy (OCA) The accuracy of a method is the ratio of correctly detected cancerous and normal cell among the total examined instance. Mathematically, accuracy can be expressed as: OCA ¼

TP þ TN  100Q TP þ TN þ FP þ FN

ð18Þ

3.6.2 Sensitivity It is defined as the measure of probability by which a method can correctly identify normal cell among total examined instance of normal cells. Mathematically, sensitivity can be expressed as: sensitivity ¼

TP  100Q TP þ FN

ð18Þ

3.6.3 Specificity It is defined as the measure of probability by which a method can correctly identify cancerous cell among total examined instance of cancerous cells. Mathematically, specificity can be expressed as: specificity ¼

TN  100Q TN þ FP

ð19Þ

3.6.4 Positive predictive value It is defined as the probability to correctly detected normal cells out of all normal lymphocyte cells by the method is can be mathematically expressed as: PPV ¼

TP  100Q TP þ FP

ð20Þ

Table 6 Brief description of experiments Experiment No

Description of Experiment

Experiment 1

The design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA- kNN classifiers and performances are given in Table 7. The design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA-PNN classifiers and performances are given in Table 8. The design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA-SSVM classifiers and performances are given in Table 9. The design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA-ANFIS classifiers and performances are given in Table 10. The design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA-SVM classifiers and performances are given in Table 11. The design of hybrid hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) by combining best performing binary classifier at each level and performances are given in Table 13.

Experiment 2 Experiment 3 Experiment 4 Experiment 5 Experiment 6

Multimed Tools Appl

3.6.5 Negative predictive value It is defined as the probability to correctly detected cancerous cells out of all leukocytes cells by the method are can be mathematically expressed as:

NPV ¼

TN  100Q TN þ FN

ð21Þ

3.7 Description of experiments In the proposed work, fastidious experiments are carried out for victimizing the cancerous and normal cell images. The concise depiction of experiments is tabulated in Table 6.

4 Result and discussion Initially, all the segmented nucleus and cytoplasm are processed for the extraction of geometrical, chromatic and statistical texture features. Geometrical features are computed by regionprops, chromatic features are computed for red, green blue band and the statistical texture descriptors are computed by using FOS, GLCM, EDGE, GLDS, NGTDM, SFM, and GLRLM statistical texture methods of nucleus and cytoplasm for each cell. The results and discussion of each experiment is described here. Experiment 1: Design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA- kNN classifier. Experiment 1 is conceded for the design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA-kNN classifier. The classification performance of each level and overall classification accuracy is reported in Table 7.

Table 7 Classification performance of hierarchical classifier using PCA-k-NN classifier Classifier

PCA- k-NN 1

PC

Confusion matrix

2 NC CC

PCA- k-NN 2

2 L1 RC

PCA- k-NN 3

5

L2 L3 Overall classification accuracy (%)

NC 58 8 L1 38 5 L2 15 1

Performance measure (%)

CC 7 57 RC 1 19 L3 4 4

BIN_ACC

Sensitivity

Specificity

PPV

NPV

88.4

87.8

89.0

89.2

87.6

90.4

88.3

95.0

97.4

79.1

79.1

93.7

50.0

78.9

80.0

80.0

FV Feature vector, PC Number of the principal components, CM Confusion Matrix, NC normal cells, CC cancerous cells, RC Remaining cells, BIN_ACC binary classification accuracy, Positive predictive value (PPV), Negative predictive value (NPV)

Multimed Tools Appl Table 8 Classification performance of hierarchical classifier using PCA-PNN classifier Classifier

PCA-PNN 1

PC

Confusion matrix

2

NC 63 3 PCA-PNN 2 5 L1 L1 38 RC 0 PCA-PNN 3 13 L2 L2 17 L3 2 Overall classification accuracy (%) NC CC

Performance measure (%)

CC 2 62 RC 1 24 L3 2 5

BIN_ACC

Sensitivity

Specificity

PPV

NPV

96.1

95.4

96.8

96.9

95.3

98.4

100

96.0

97.4

100

91.6

100.0

71.4

89.4

100.0

92.3

Experiment 2: Design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA- PNN classifier.

Experiment 2 is conceded for the design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA-PNN classifier. The classification performance of each level and overall classification accuracy is reported in Table 8. Experiment 3: Design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA- SSVM classifier.

Experiment 3 is conceded for the design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA-SSVM classifier. The classification performance of each level and overall classification accuracy is reported in Table 9. Experiment 4: Design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA- ANFIS classifier.

Table 9 Classification performance of hierarchical classifier using PCA-SSVM classifier Classifier

PCA-SSVM 1

PC

Confusion matrix

15 NC CC

PCA-SSVM 2

13 L1 RC

PCA-SSVM 3

12

L2 L3 Overall classification accuracy (%)

NC 60 7 L1 37 3 L2 16 1

Performance measure (%)

CC 5 58 RC 2 20 L3 3 4

BIN_ACC

Sensitivity

Specificity

PPV

NPV

90.7

89.5

92.0

92.3

89.2

90.4

90.2

90.9

94.8

83.3

83.3

94.1

57.1

84.2

80.0

83.8

Multimed Tools Appl Table 10 Classification performance of hierarchical classifier using PCA-ANFIS classifier Classifier

PCA-ANFIS 1

PC

Confusion Matrix

15 NC CC

PCA-ANFIS 2

12 L1 RC

PCA-ANFIS 3

5

L2 L3 Overall classification accuracy (%)

NC 62 4 L1 37 3 L2 18 0

Performance measure (%) BIN_ACC

Sensitivity

Specificity

PPV

NPV

94.6

93.9

95.3

95.3

93.8

92.0

92.5

91.3

94.8

87.5

95.8

100.0

83.3

94.7

100.0

CC 3 61 RC 2 21 L3 1 5

90.0

Experiment 4 is conceded for the design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA-PNN classifier. The classification performance of each level and overall classification accuracy is reported in Table 10. Experiment 5: Design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA- SVM classifier.

Experiment 5 is conceded for the design of hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using PCA-PNN classifier. The classification performance of each level and overall classification accuracy is reported in Table 11. The overall classification accuracy is calculated by adding the number of misclassifications obtained at each level of the hierarchical classifiers. Therefore PCA-SVM classifiers yields minimum i.e. a total of 7 misclassifications consisting of 1, 1, 3 and 2 misclassifications for PCA-SVM1, PCA-SVM2 and PCA-SVM3 classifiers respectively, thus overall classification accuracy for PCA-SVM based hierarchical classifiers is 94.6% i.e. {(130–7) / 130} × 100 = {(123 / 130) × 100} = 94.6%.

Table 11 Classification performance of hierarchical classifier using PCA- SVM classifier Classifier

PCA-SVM 1

PC

Confusion matrix

6

NC 65 1 PCA-SVM 2 13 L1 L1 38 RC 3 PCA-SVM 3 9 L2 L2 17 L3 0 Overall classification accuracy (%) NC CC

Performance measure (%)

CC 0 64 RC 1 21 L3 2 5

BIN_ACC

Sensitivity

Specificity

PPV

NPV

99.2

98.4

100.0

100.0

98.4

93.6

92.6

95.4

97.4

87.5

91.6

100.0

71.4

89.4

100.0

94.6

Multimed Tools Appl Table 12 Comparative analysis of performances for designed hierarchical classifiers using various experiments Experiment

BIN_ACCL1 (%)

BIN_ACC L2 (%)

BIN_ACC L3 (%)

OCA (%)

MI

PCA-k-NN PCA-PNN PCA-SSVM PCA-ANFIS PCA- SVM

88.4 96.1 90.7 94.6 99.2

90.4 98.4 90.4 92.0 93.6

79.1 91.6 83.3 95.8 91.6

80.0 92.3 83.8 90.0 94.6

26 10 21 13 7

BIN_ACC Accuracy of a binary classifier, L1 level one, L2 level two, L3 level three, OCA Overall classification accuracy, MI Misclassified instances

By visualizing the performance of individual binary classifiers of PCA-kNN, PCA-PNN, PCA-SSVM, PCA-ANFIS and PCA-SVM based hierarchical classifiers (shown in Tables 6, 7, 8, 9 and 10), few facts are observed: (i) In classification module one the classification between NC versus CC the maximum accuracy of 99.2% is obtained by using PCA-SVM1 classifier in comparison with 88.4%, 96.1%, 90.7% and 94.6% as obtained by using PCA-kNN1, PCA-PNN1, PCA-SSVM1, and PCA- ANFIS1 classifiers. (ii) In classification module two, further classification of CC into FAB classes in step wise order i.e. L1 versus RC the maximum accuracy of 98.4% is obtained by using PCA-PNN2 classifier in comparison with 90.4%, 90.4%, 92.0% and 93.6% as obtained by using PCAkNN2, PCA-SSVM2, PCA-ANFIS2 and PCA-SVM2 classifiers respectively. (iii) Further the classification of RC into L2 versus L3 class the maximum accuracy of 95.8% is obtained by using PCA- ANFIS3 classifier in comparison with 79.1%, 91.6%, 83.3% and 91.6% as obtained by using PCA- kNN 3, PCA-PNN3, PCA-SSVM3 and PCASVM3 classifiers.

Fig. 9 Architecture of hybrid hierarchical classifier

Multimed Tools Appl Table 13 Classification performance of the hybrid hierarchical classifier Classifier

PC

PCA-SVM 1

Confusion Matrix

6

NC 65 1 L1 38 0 L2 18 0

NC CC PCA-PNN 2

5 L1 RC

PCA-ANFIS 3

5

L2 L3 Overall classification accuracy (%)

Performance measure (%) BIN_ACC

Sensitivity

Specificity

PPV

NPV

99.2

98.4

100.0

100.0

98.4

98.4

100

96.0

97.4

100

95.8

100.0

83.3

94.7

100.0

CC 0 64 RC 1 24 L3 1 5

97.6

Comparative analysis The comparative analysis of performances for various hierarchical classifiers is given in Table 12. From the Table 12 it has been observed that the maximum classification accuracy at level one is 99.2% using PCA-SVM classifiers at pc value 6. Similarly 98.4% and 95.8% of classification accuracy at level two, level three is obtained using PCA-PNN and PCAANFIS classifiers at pc value 5 and 5 respectively. Further, the total misclassified instances are 8 using PCA-SVM classifiers which are less with respect to misclassifications yielded by PCA-kNN, PCA-PNN, PCA-SSVM and PCA-ANFIS classifiers. Experiment 6: Design of hybrid hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) by combining best performing binary classifier at each level. Experiment 6 is conceded for the design of hybrid hierarchical classifier for NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) using best performing binary classifier at each level. The architecture of the hybrid hierarchical classifiers is shown in Fig. 9. The classification performance of each level and overall classification accuracy is reported in Table 13. From the Table 13, it has been observed that the maximum classification accuracy increases from 94.6 to 97.6% after arranging the best binary classifier at each level in hierarchical order. The Table 14 Database description for ALL classification. [25] Image acquisition

ALL-IDB1 ALL-IDB2 ALL-IDB2 FAB classes No. of images Dataset bifurcation Training dataset Training dataset

Camera Colour depth Magnification power Image format

Canon powerShot G5 24 bit 300 to 500 JPG

Images

Resolution

Elements

Candidates

109 260

2592 × 1944 257 × 257

39,000 260

510 130

L1 79

L2 39

L3 10

40 39

20 19

5 5

Multimed Tools Appl

total number of misclassified instances decreases from 7 to 3 with a comparison of hierarchical classifiers (PCA-kNN, PCA-PNN, PCA-SSVM, PCA-ANFIS and PCA-SVM). The misclassified instances consisting of one at level one, one at level two and one at level three. The various classifiers have been designed in hierarchical order to classify NC/CC cell and FAB classification of acute lymphoblastic leukaemia (ALL) and the hybrid hierarchical classifier is designed by arranging the best binary classifier at each level in hierarchical order. The promising results obtained by the hybrid hierarchical indicate that the proposed CAD system shall be used in the pathological lab to detect the cancerous cell and further classify it according to FAB classes (Table 14).

5 Conclusion A new automated hybrid hierarchical classification method for 2-class (Normal cells and Cancerous or ALL cells) and 3-class (L1, L2, and L3) has been proposed. The proposed method is capable to classify the cancerous lymphoblast cells from the normal one and properly discriminate between the FAB classes of ALL (L1, L2, and L3) with enormous overall classification accuracy of 97.6%, by using hybrid hierarchical classifier with (i) finding true edges of nucleus and cytoplasm, (ii) correct classification of cancerous malignant lymphoblast cells and normal cells, (iii) correct classification of FAB classes of ALL (L1, L2, and L3). The characterisation performance of the system is evaluated by multiple classifiers (PCA-kNN, PCA-PNN, PCA-SSVM, PCA-ANFIS, and PCA-SVM). The results reveal that the proposed method performs significantly well in distinguishing between the normal and cancerous cell with the accuracy of 99.2%. Further Cancerous cell is classified as per FAB classes with the accuracy of 98.4% for L1 and remaining class and 95.8% is achieved for L2 versus L3 class. The major contribution of present work is to improve earlier classification methods by employing multiple classification problems into a single system. Classification result indicates that the localization of malignant white cells (lymphoblast) is achievable and offer remarkable classification accuracy for categorizing FAB classes of ALL. The method is efficient and robust to strongly support the medical application by prior identification of lymphoblast cancerous cell. The promising results obtained by the hybrid hierarchical classifier point towards the proposed CAD system can be utilized in the field haematology for the early detection of ALL. In spite of the great outcomes acquired with present work, further expansions can be made to the proposed scheme. Specifically can enhance strength and exactness of the classification task, which is a critical issue particularly for microscopic images of blood cells. The investigation of other features could be expanded likewise for the interpretation of other medicinal issues, including the study of other type of haematological disorders like AML (acute myeloid leukemia) and other kind of leukemias that can influence the kind of cell. Moreover, there is an intension to apply the proposed algorithm on a bigger database.

References 1. Aimi AN, Mashor MY, Rosline H (2013) Classification of acute leukaemia cells using multilayer perceptron and simplified fuzzy ARTMAP neural networks 2. American Cancer Society (2015) Facts Spring

Multimed Tools Appl 3. Amin MM, Kermani S, Talebi A, Oghli MG (2015) Recognition of acute lymphoblastic leukemia cells in microscopic images using K-means clustering and support vector machine classifier. Journal of medical signals and sensors 5(1):49 4. Bain BJ (2008) A beginner's guide to blood cells. Wiley 5. Banday SA, Mir AH (2016) Statistical textural feature and deformable model based brain tumor segmentation and volume estimation. Multimedia Tools Appl 6. Bhattacharjee R, Saini LM (2015) Robust technique for the detection of acute lymphoblastic leukemia. Proc. of Int Conf. on Power, Communication and Information Technology Conference. IEEE, In, pp 657– 662 7. Chandy DA, Johnson JS, Selvan SE (2014) Texture feature extraction using gray level statistical matrix for content-based mammogram retrieval. Multimedia Tools Application 72:2011 8. Chang CC, Lin CJ (2012) LIBSVM, A library of support vector machines Available at http://www.csie.ntu. edu.tw/~cjlin/libsvm 9. Choi YH, Tak YS, Rho S (2013) Skin feature extraction and processing model for statistical skin age estimation. Multimedia Tools Application 64:227 10. Chris A, Mulyawan B (2012) A combination of feature selection and co-occurrence matrix methods for leukocyte recognition system. J Softw Eng Appl 5:101 11. Dash JK, Mukhopadhyay S, Gupta RD (2016) Multiple classifier system using classification confidence for texture classification. Multimedia Tools Appl 12. David HA (1981) Order statistics. Willey, New York 13. ElDahshan K, Youssef M, Masameer E, Mustafa MA (2015) An efficient implementation of acute lymphoblastic leukemia images segmentation on the FPGA. Advances in Image and Video Processing 3(3):8 14. Gonzalez RC, Woods RE, Eddins SL (2004) Digital image using Matlab processing. Lexington, Person Prentice Hall 15. Gorunescu M et al (2005) A heuristic approach in hepatic cancer diagnosis using a probabilistic neural network based model. Proc. of Int Conf on Applied Stochastic Models and Data Analysis, France, In, pp 1016–1025 16. Halim NHA, Mashor MY, Hassan R (2011) Automatic blasts counting for acute leukemia based on blood samples. International Journal of Research and Reviews in Computer Science 2(4):971–976 17. Han ZY, Gu DH, Wu QE (2016) feature extraction for color images. In: Electronics, Communications and Networks V 215–221 18. Haralick RM, Shanmugam K, Dinstein I (1973) Textural features for image classification. IEEE T Syst Man Cyb 3(6):610–621 19. Hinge CS et al (2013) Classification of RBC and WBC in peripheral blood smear using KNN. Indian Journal of Research 2:2250–1991 20. Hu H, Li Y, Liu M (2014) Classification of defects in steel strip surface based on multiclass support vector machine. Multimedia Tools Application 69:199 21. Khan S, Hussain M, Aboalsamh H (2015) A comparison of different Gabor feature extraction approaches for mass classification in mammography. Multimedia Tools Application 76:33–57 22. Kher R, Pawar T, Thakar V, Shah H (2015) Physical activities recognition from ambulatory ECG signals using neuro-fuzzy classifiers and support vector machines. Journal of medical engineering & technology 39(2):138–152 23. Kovalev VA, Grigoriev AY, Ahn H (1996) Robust recognition of white blood cell images. In: Proc. of Int Conf on Pattern Recognition 4. IEEE, pp 371–375 24. Kumar I, Bhadauria HS, Virmani J, Rawat J (2014) Reduction of speckle noise from medical images using principal component analysis image fusion. Proc. of Int Conf. on Industrial and Information Systems. IEEE, In, pp 1–6 25. Labati RD, Piuri V (2045-2048) Scotti F (2011) all-IDB: the acute lymphoblastic leukemia image database for image processing. In, Proc. of Int Conf. on Image Processing. IEEE, pp 26. Lee YJ, Mangasarian OL (2015) SSVM toolbox Available at: http://research.cs.wisc.edu/dmi/svm/ssvm 27. Leukaemia and Lymphoma Society (2015) Blood and marrow stem cell transplantation 28. Madhloom HT, Kareem SA, Ariffin H, Zaidan AA, Alanazi HO, Zaidan BB (2010) An automated white blood cell nucleus localization and segmentation using image arithmetic and automatic threshold. J Appl Sci 10:959–966 29. Madhloom HT, Kareem SA, Ariffin H (2012) A robust feature extraction and selection method for the recognition of lymphocytes versus acute lymphoblastic leukemia. Proc. of Int Conf. on Advanced Computer Science Applications and Technologies. IEEE, In, pp 330–335 30. Manth N, Virmani J, Kumar V, Kalra N, Khandelwal N (2015) Despeckle filtering: performance evaluation for malignant focal hepatic lesions. In: Proc. of Int Conf. On Computing for Sustainable Global Development 1897–1902, IEEE 31. Mohapatra S, Patra D (2010a) Automated cell nucleus segmentation and acute leukemia detection in blood microscopic images. Proc. of Int Conf. Systems in Medicine and Biology. IEEE, In, pp 49–54 32. Mohapatra S, Patra D (2010b) Automated leukemia detection using hausdorff dimension in blood microscopic images. Proc. of Int Conf. on Emerging Trends in Robotics and Communication Technologies. IEEE, In, pp 64–68

Multimed Tools Appl 33. Mohapatra S, Patra D, Satpathi S (2010) Image analysis of blood microscopic images for acute leukemia detection. Proc. of Int Conf on Electronics, Control & Robotics. IEEE, In, pp 215–219 34. Mohapatra S, Samanta SS, Patra D, Satpathi S (2011) Fuzzy based blood image segmentation for automated leukemia detection. Proc. of Int Conf. on Devices and Communications. IEEE, In, pp 1–5 35. Mohapatra S, Patra D, Kumar S, Satpathy S (2012) Lymphocyte image segmentation using functional link neural architecture for acute leukemia detection. Biomed Eng Lett 2(2):100–110 36. Mohapatra S, Patra D, Satpathy S (2014a) An ensemble classifier system for early diagnosis of acute lymphoblastic leukemia in blood microscopic images. Neural Comput & Applic 24(7–8): 1887–1904 37. Mohapatra S, Patra D, Satpathy S, Jena RK, Sethy S (2014b) Automated morphometric classification of acute lymphoblastic leukaemia in blood microscopic images using an ensemble of classifiers. Computer Methods in Biomechanics and Biomedical Engineering, Imaging & Visualization 4(1):3–16 38. MoradiAmin M, Memari A, Samadzadehaghdam N, Kermani S, Talebi A (2015) Enhanced recognition of acute lymphoblastic leukemia cells in microscopic images based on feature reduction using principle component analysis. Frontiers in Biomedical Technologies 2(3):128–136 39. MoradiAmin M, Memari A, Samadzadehaghdam N, Kermani S, Talebi A (2016) Computer aided detection and classification of acute lymphoblastic leukemia cell subtypes based on microscopic image analysis. Microsc Res Tech 79:908–916 40. Nasir AA, Mashor MY, Rosline H (2011) Detection of acute leukaemia cells using variety of features and neural networks. Proc. of Int Conf. on Biomedical Engineering, Springer, Berlin, In, pp 40–46 41. Neoh SC, Srisukkham W, Zhang L, Todryk S, Greystoke B, Lim CP, Hossain MA, Aslam N (2015) An intelligent decision support system for leukaemia diagnosis using microscopic blood images. Scientific reports 5:14938 42. Pedreira CE, Macrini L, Land MG, Costa ES (2009) New decision support tool for treatment intensity choice in childhood acute lymphoblastic leukemia. IEEE Trans Inf Technol Biomed 13(3):284–290 43. Putzu L, Di Ruberto C (2013) White blood cells identification and classification from leukemic blood image. Proc. of Int Conf. on bioinformatics and biomedical engineering, In, pp 99–106 44. Putzu L, Caocci G, Di Ruberto C (2014a) Leucocyte classification for leukaemia detection using image processing techniques. Artif Intell Med 62(3):179–191 45. Putzu L, Caocci G, Di Ruberto C (2014b) Leucocyte classification for leukaemia detection using image processing techniques. Artif Intell Med 62(3):179–191 46. Rawat J, Singh A, Bhadauria HS (2014a) An approach for leukocytes nuclei segmentation based on image fusion. Proc. of Int Symposium on Signal Processing and Information Technology. IEEE, In, pp 456–461 47. Rawat J, Singh A, Bhadauria HS, Kumar I (2014b) Comparative analysis of segmentation algorithms for leukocyte extraction in the acute lymphoblastic leukemia images. Proc. of Int Conf. on Parallel, Distributed and Grid Computing. IEEE, In, pp 245–250 48. Rawat J, Singh A, Bhadauria HS, Virmani J (2015) Computer aided diagnostic system for detection of leukemia using microscopic images. Procedia Computer Science 70:748–756 49. Sabino DMU, da Fontoura CL, Rizzatti EG, Zago MA (2004) A texture approach to leukocyte recognition. Real-Time Imaging 10(4):205–216 50. Saipullah KM, Kim DH (2012) A robust texture feature extraction using the localized angular phase. Multimedia Tools Application 59:717 51. Saraswat M, Arya KV (2014) Automated microscopic image analysis for leukocytes identification: a survey. Micron 65:20–33 52. Scotti F (2005) Automatic morphological analysis for acute leukemia identification in peripheral blood microscope images. Proc. of Int Conf. on Computational Intelligence for Measurement Systems and Applications, In, pp 96–101 53. Scotti F (2006) Robust segmentation and measurements techniques of white cells in blood microscope images. Proc. of Int Conf on Instrumentation and Measurement Technology. IEEE, In, pp 43–48 54. Sharma K, Virmani J (2016) Classification of renal diseases using first order and higher order statistics. In: Proc. of Conf. on Computing for Sustainable Global Development (INDIACom). IEEE, 425–430 55. Singh G, Bathla G, Kaur S (2016) Design of New Architecture to detect leukemia cancer from medical images. Int J Appl Eng Res 11(10):7087–7094

Multimed Tools Appl 56. Singhal V, Singh P (2015) Correlation based feature selection for diagnosis of acute lymphoblastic leukemia. Proc. of Int Symposium on Women in Computing and Informatics. ACM, In 57. Singhal V, Singh P (2016) Texture features for the detection of acute lymphoblastic leukemia. Proc. of Int. Conf. on ICT for Sustainable Development. Springer, In, pp 535–543 58. Sinha N, Ramakrishnan AG (2003) Automation of differential blood count. In: Proc. of TENCON 2. IEEE, pp 547–551 59. Virmani J, Kumar V, Kalra N, Khandelwal N (2013a) Characterization of primary and secondary malignant liver lesions from B-mode ultrasound. J Digit Imaging 26(6):1058–1070 60. Virmani J, Kumar V, Kalra N, Khandelwal N (2013b) PCA-SVM based CAD system for focal liver lesions using B-mode ultrasound images. Defence Sci J 63(5):478–486 61. Virmani J, Kumar V, Kalra N, Khandelwal N (2013c) SVM-based characterisation of liver cirrhosis by singular value decomposition of GLCM matrix. International Journal of Artificial Intelligence and Soft Computing 3(3):276–296 62. Virmani J, Kumar V, Kalra N, Khandelwal N (2013d) SVM-based characterization of liver ultrasound images using wavelet packet texture descriptors. J Digit Imaging 26(3):530–543 63. Virmani J, Kumar V, Kalra N, Khandelwal N (2013e) A comparative study of computer-aided classification systems for focal hepatic lesions from B-mode ultrasound. J Med Eng Technol 37(4):292–306 64. Virmani J, Kumar V, Kalra N, Khandelwal N (2014) Neural network ensemble based CAD system for focal liver lesions from B-mode ultrasound. J Digit Imaging 27(4):520–537 65. Viswanathan P (2015) Fuzzy C means detection of leukemia based on morphological contour segmentation. Procedia Computer Science 58:84–90 66. Wong HS, Cheung KKT, Chiu CI (2007) Hierarchical multi-classifier system design based on evolutionary computation technique. Multimedia Tools Application 33:91 67. Yang M, Kpalma K, Ronsin J (2008) A survey of shape feature extraction techniques. Pattern Recogn:43–90 68. Zhang L, Wang QG, Qi JP (2006) Processing Technology in Microscopic Images of cancer cells in pleural fluid based on fuzzy edge detection method. J Phys 48(1):329

Jyoti Rawat received her diploma (Hons.) in Information Technology from Govt. Polytechnic Srinagar; B Tech (Hons.) in Information Technology from Rajasthan Technical University, Kota, Rajasthan, in 2010 and M Tech in Computer Science and Engineering from Graphic Era University, Dehradun, Uttarakhand, India in 2013. Presently, She is pursuing PhD in Computer Science and Engineering from G B Pant Engineering College, Pauri Garhwal, Uttarakhand and her research area is Medical image processing. Her research interests include application of digital image processing, machine learning and soft computing techniques for analysis of medical images.

Multimed Tools Appl

Annapurna Singh received her M.Tech from Banasthali Vidyapeeth, Rajasthan, India in 2003; she did her PhD from Uttarakhand Technical University Dehradun in Computer Science Engineering in 2013. From 2003 to till date she worked as lecturer in various Engineering Colleges of India and presently working as an Assistant Professor in Computer Science and Engineering department of G.B. Pant Engineering College, India. Her research interests include application of machine learning, Digital Image and Digital Signal.

H. S. Bhadauria received his B Tech in Computer science and Engineering from Aligarh Muslim University, Aligarh in 1999, and M. Tech. in Electronics Engineering from Aligarh Muslim University, Aligarh in 2004. He received his PhD on Detection and Segmentation of Brain Hemorrhage using CT images from Biomedical Signal and Image Processing Laboratory, Indian Institute of Technology - Roorkee in 2013. During his PhD he worked on enhancing the detection and segmentation of brain hemorrhage using CT imaging modality. He served in academia for 12 years. He is presently serving as Assistant Professor at Govind Ballabh Pant Engineering college, Pauri Garhwal, Uttarakhand. He is a life member of Institute of Engineers (IEI), India. He has published more than 60 research papers in International and National Journals and Conferences. His areas of Interest are Digital Image and Digital Signal Processing.

Multimed Tools Appl

Jitendra Virmani received his B Tech (Hons.) in Instrumentation Engineering from Sant Longowal Institute of Engineering and Technology, Punjab in 1999 and M Tech in Electrical Engineering with specialization in Measurement and Instrumentation Engineering from Indian Institute of Technology, Roorkee in 2006. He received his PhD on Analysis and Classification of B-Mode Liver Ultrasound Images from Biomedical Signal and Image Processing Laboratory, Indian Institute of Technology - Roorkee in 2014. During his PhD he worked on enhancing the potential of most commonly available B-Mode Ultrasound imaging modality for differential diagnosis between atypical cases of focal liver lesions. He served in academia in various reputed organizations like Jaypee University of Information Technology- Solan, H.P, India and Thapar University- Patiala, Punjab, India for 13 years before joining the CSIR-CSIO, Central Scientific Industrial Research-Central scientific instruments organization, Chandigarh, India during Aug 2016. He is a life member of Institute of Engineers (IEI), India. His research interests include application of machine learning and soft computing techniques for analysis of medical images.

J. S. Devgun received his MBBS degree from GSVM medical college, Kanpur, India and his MD Degree in the field of Radio-diagnosis from MLB medical college, Jhansi, India. Thereafter he pursued Registrarship in the same field in M. M. Institute of Medical Sciences & Research and hospital, Solan, India. He is a renowned researcher in his field with more than 15 publications in journals of international and national repute.