Classiﬁcation of Normal versus Leukemic Cells with Data Augmentation

and Convolutional Neural Networks

Jos

e Elwyslan Maur

ıcio de Oliveira

and Daniel Oliveira Dantas

Departamento de Computac¸

ao, Universidade Federal de Sergipe, S

ao Crist

ao, SE, Brazil

Keywords:

Leukemia Classiﬁcation, Acute Lymphoblastic Leukemia.

Abstract:

Acute lymphoblastic leukemia is the most common childhood leukemia. It is an aggressive cancer type and

causes various health problems. Diagnosis depends on manual microscopic analysis of blood samples by

expert hematologists and pathologists. To assist these professionals, image processing and pattern recognition

techniques can be used. This work proposes simple modiﬁcations to standard neural network architectures

to achieve high performance in the malignant leukocyte classiﬁcation problem. The tested architectures were

VGG16, VGG19 and Xception. Data augmentation was employed to balance the Training and Validation

sets. Transformations such as mirroring, rotation, blurring, shearing, and addition of salt and pepper noise

were used. The proposed method achieved an F1-score of 92.60%, the highest one when compared to other

participants’ published results and eighth position when compared to the weighted F1-score provided by the

competition leaderboard.

1 INTRODUCTION

Blood is a connective tissue that ﬂows within the

blood vessels of animals that have a closed circula-

tory system. In hematology, the branch of medicine

concerned with the study of blood, changes in the

shape and function of leukocytes are called leukocyte

abnormalities, and leukemia is one of those abnor-

malities. These abnormalities are deﬁned by the ac-

cumulation of myeloblasts (immature granulocytes)

or lymphoblasts (immature lymphocytes) in the bone

marrow and peripheral blood. Leukemia can occur in

two forms: acute or chronic. Acute leukemia is the

most aggressive since it evolves rapidly and presents

its symptoms more intensely than the chronic version.

Acute Lymphoblastic Leukemia (ALL), the main

object of this study, occurs when a large number of

lymphoblasts accumulate in the bone marrow and pe-

ripheral blood. These immature lymphocytes do not

differentiate into their mature forms, causing health

problems such as infections or cancer. ALL is the

most common childhood leukemia. The highest oc-

currence of ALL being in children between 3 and 7

years old with 75% of diagnoses occurring before the

age of 6. Of these diagnoses, 85% are from leukemias

https://orcid.org/0000-0002-0282-0921

https://orcid.org/0000-0002-0142-891X

that affect B type lymphocytes (ALL-B) (Hoffbrand

and Moss, 2013).

Manual microscopic analysis of blood samples is

the primary method for analyzing lymphocytes ex-

tracted from patients with leukemia. As a result, the

classiﬁcation of healthy and malignant lymphocytes

highly depends upon the expertise of the hematolo-

gists and pathologists in recognizing the two classes.

To assist these professionals in live blood analysis,

image processing and pattern recognition techniques

have been extensively used to produce Computer-

Aided Diagnosis (CADx) systems. These systems

aim to increase the accuracy of lymphocyte classiﬁ-

cation (Mishra et al., 2019; Moshavash et al., 2018).

The Acute Lymphoblastic Leukemia Image

Database (ALL-IDB) (Labati et al., 2011; DI-UNIMI,

2020) for Image Processing is a public and free

dataset of microscopic images of blood samples for

the evaluation of segmentation and image classiﬁca-

tion algorithms focusing on ALL. The ALL-IDB ini-

tiative provides two different datasets: ALL-IDB1,

which consists of 108 blood smear images collected

from healthy and leukemic patients, containing 510

single leukocytes; and ALL-IDB2 which is a col-

lection of the cropped areas of interest of normal

and malignant leukocytes that belong to the ALL-

IDB1 dataset. Samples of both ALL-IDB datasets are

shown in Figure 1.

Maurício de Oliveira, J. and Dantas, D.

Classiﬁcation of Normal versus Leukemic Cells with Data Augmentation and Convolutional Neural Networks.

DOI: 10.5220/0010257406850692

In Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2021) - Volume 4: VISAPP, pages

685-692

ISBN: 978-989-758-488-6

685

Figure 1: ALL-IDB image samples (DI-UNIMI, 2020): (a)

ALL-IDB1 and (b) ALL-IDB2.

Many works in the literature use small datasets.

Usage of the dataset ALL-IDB, which contains

only 510 lymphocytes, was commonplace, and

many feature extraction based approaches were pro-

posed (Putzu et al., 2014; MoradiAmin et al., 2016;

Mishra et al., 2016; Rawat et al., 2017; Moshavash

et al., 2018). However, most current works use con-

volutional neural networks (CNN) to address the lym-

phocyte classiﬁcation task.

Shaﬁque and Tehsin (Shaﬁque and Tehsin, 2018)

deployed a pretrained AlexNet for detection and clas-

siﬁcation of ALL lymphocytes. ALL-IDB2 dataset

was used in combination with data augmentation to

increase the number of training samples by apply-

ing image rotation and mirroring in the source im-

ages. The training data was increased from 260 im-

ages to 760, where 500 are malignant samples and

260 healthy samples. The augmented dataset was di-

vided into training data and test data in a 6:4 ratio.

The pretrained AlexNet achieved an overall accuracy

of 99.50%.

Rehman et al. (Rehman et al., 2018) also propose

a deep learning solution using Alexnet and transfer

learning. In his work, only the top layers of Alexnet

were modiﬁed and ﬁne-tuned. ALL-IDB was used

without data augmentation. The test set contained

330 images, and the network achieved an accuracy of

97.78%.

Ahmed et al. (Ahmed et al., 2019) conducted

some experiments with ALL-IDB, in one of which

they had develop their own CNN architecture. Data

augmentation was used to balance and expand the

dataset. The training was performed with 980 sam-

ples, and an additional 245 samples were used as test,

with both sets being balanced. His CNN achieved an

accuracy of 88.5%.

Recent initiatives have released large datasets to

train and evaluate leukemia cell classiﬁers. One is the

CNMC-2019 dataset created by the SBILab (SBILab,

2020) research team. This dataset is used in this work

and is detailed in Subsection 2.1.

Mourya et al. (Mourya et al., 2018) (SBILab

members) introduced a deep learning framework for

classifying malignant and healthy lymphocytes that

combine discrete cosine transform (DCT) features

and convolutional neural networks (CNN) in a hybrid

architecture called Leukonet. They have prepared a

dataset of 9211 cancer cells from 65 subjects and

4528 healthy cells from 52 subjects, which together

composed the training data of 13739 cells, divided

into four sets for cross-validation. The test data used

to evaluate Leukonet consist of 312 cancer cells and

324 healthy cells. Its hybrid architecture achieved an

accuracy of 89.70% and an F1-score of 91.95% for

cancer cell class.

In 2019 the SBILab hosted a competition in which

participants are to make use of the C-NMC 2019

Dataset to propose classiﬁcation methods for ALL-B

cells into healthy or malignant. At the end of the com-

petition, winning approaches were published (Gupta

and Gupta, 2019). All submitted solutions were based

on convolutional neural networks, and many had large

and complex architectures. We show throughout this

paper that it is possible to achieve a high score in this

challenge using slightly modiﬁed traditional architec-

tures and standard training methods.

This article is organized as follows: Section 2 de-

scribes the methodology used in this work; Section 3

presents the obtained results, and; the conclusions are

given in Section 4.

2 METHODOLOGY

In this work, images of healthy and malignant lym-

phocytes from C-NMC 2019 Dataset were used for

training variations of the Xception and VGGNet ar-

chitectures, in order to create a classiﬁer capable of

distinguishing the two cell types. These architectures

were ﬁne-tuned so that the models achieved the high-

est validation accuracy possible. In the end, the best

models of each architecture had their performances

evaluated in the Test set.

The implementation of this methodology is pub-

licly available

and was coded in Python using Ten-

sorﬂow, Keras, Numpy, SciPy and OpenCV.

https://github.com/Elwyslan/ISBI-2019-Cells-

Classiﬁcation/tree/master/classiﬁers/10.ConvolutionalNets

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

686

Figure 2: C-NMC 2019 Dataset samples. (a) and (c) are ma-

lignant lymphocytes, (b) and (d) are healthy lymphocytes.

2.1 C-NMC 2019 Dataset

The SBILab (SBILab, 2020) research team was

responsible for creating and publishing the image

dataset used in this work. They performed all the

steps related to image preprocessing, image enhance-

ment, lymphocyte segmentation, and stain normaliza-

tion using standard image processing techniques and

inhouse methods (Duggal et al., 2016; Gupta et al.,

2017; Duggal et al., 2017).

The C-NMC 2019 Dataset, publicly avail-

able (Mourya et al., 2019), consists of 15114 lym-

phocyte images collected from 118 subjects and split

into three folders with names: “C-NMC training data”

containing 10661 cells, 7272 malignant cells from 47

subjects and 3389 healthy cells from 26 subjects; “C-

NMC test preliminary phase data” containing 1867

cells, 1219 malignant cells from 13 subjects and 648

healthy cells from 15 subjects, and “C-NMC test ﬁ-

nal phase data” containing 2586 unlabeled cells from

17 subjects. Within these folders, there are single cell

images of malignant and healthy lymphocytes previ-

ously labeled by expert oncologists.

Cells had been dyed using Jenner-Giemsa stain

technique (Marzahl et al., 2019), and the images were

stain-normalized before the segmentation. The blood

smear image is a bitmap RGB with 24 bits color

depth and size of 2560x1920 pixels. The lymphocytes

were segmented from each of the blood smear images

and placed in the center of individual images of size

450×450 pixels with a black background. Figure 2

Figure 3: Examples of augmented images: (a) source im-

age; (b) vertical and horizontal mirroring; (c) clockwise

rotation by 60

◦

; (d) Gaussian blur with 17×17 kernel; (e)

shearing with a factor of 0.3, and; (f) salt and pepper noise.

Table 1: Number of samples in Training, Validation and

Test sets.

Before

Data Augmentation

After

Data Augmentation

Malignant Healthy Malignant Healthy

Training

4364 2034 10000 10000

Validation

2181 1016 5000 5000

Test

727 339 N/A N/A

shows samples of malignant and healthy lymphocytes

of the C-NMC 2019 Dataset.

The unlabeled data in the C-NMC test ﬁnal phase

data is used to evaluate the medical image classiﬁca-

tion challenge entitled “C-NMC challenge: Classiﬁ-

cation of Normal versus Malignant Cells in B-ALL

White Blood Cancer Microscopic Images” organized

by SBILab (SBILab, 2019a). The participants in this

challenge can evaluate their results on C-NMC test

ﬁnal phase data by submitting it to the leaderboard

hosted on competition website.

In this work, the “C-NMC training data” was used,

whose images were randomly split into Training, Val-

idation, and Test sets in a 6:3:1 ratio. The division

of the 7272 malignant lymphocytes was as follows:

4364 were for the Training set, 2181 for the Valida-

tion set, and 727 for the Test set. Regarding the 3389

healthy lymphocytes: 2034 were for the Training set,

1016 for the Validation set and 339 for the Test set.

The split is shown in Table 1.

2.2 Data Augmentation

The original dataset was unbalanced and, for that rea-

son, data augmentation was employed to balance the

Training and Validation sets. This technique was not

applied to the Test set. Standard image transforma-

tion techniques were used, such as mirroring, rotation,

Classiﬁcation of Normal versus Leukemic Cells with Data Augmentation and Convolutional Neural Networks

687

and Gaussian blurring, to produce the augmented im-

ages. We also attempted different augmentation tech-

niques, and it was noticed that shearing and addition

of salt and pepper noise resulted in a better train-

ing performance and model accuracy. An example

of these techniques applied to a random image from

the dataset can be seen in Figure 3. The augmented

Training set had 20,000 samples, and the Validation

set has 10,000 samples, as shown in Table 1.

2.3 Convolutional Neural Network

Classiﬁer

The VGG16, VGG19, and Xception architectures

were chosen to build the classiﬁers. The Xcep-

tion (Chollet, 2017) and VGGNet (Simonyan and Zis-

serman, 2014) were the best qualiﬁed CNN archi-

tectures presented in the ImageNet Large Scale Vi-

sual Recognition Challenge (ILSVRC) (Russakovsky

et al., 2015; Dhillon and Verma, 2019).

The VGGNet, proposed by the Visual Geometry

Group (VGG) from Oxford University, has six dif-

ferent convolutional network conﬁgurations by the

names of: VGG11, VGG11-LRN, VGG13, VGG16

(Conv1), VGG16, and VGG19. Each of these conﬁg-

urations has the number of convolutional layers equal

to the number associated with its name. In ILSVRC,

VGG16 and VGG19 achieved the highest accuracy.

The VGG16 and VGG19 top layers consist of a

global max pooling layer followed by two fully con-

nected layers with 4096 neurons using ReLU activa-

tion function. The output layer is made of 1000 neu-

rons using softmax function. In our work, these layers

were replaced by a global average pooling layer fol-

lowed by two fully connected layers with 512 neurons

using ReLU, then linked to a prediction layer with two

neurons using softmax function.

The Xception extends the concept of performing

several convolutions with different ﬁlter sizes from

Inception’s module by using the concept of depthwise

separable convolutions. This architecture is com-

posed of 36 depthwise separable convolution layers,

structured in 14 modules. The modules have residual

connections to each other, except for the ﬁrst and last

modules (Chollet, 2017).

The Xception top layers consist of a global av-

erage pooling layer which produces a 1×2048 vec-

tor. In the paper that describes the architecture, Chol-

let (Chollet, 2017) does not specify any fully con-

nected or prediction layer, therefore we decided to

place one fully connected layer with 2048 neurons us-

ing ReLU linked to 2 neurons using softmax.

The ﬁrst paper to propose Global Average Pooling

(GAP) layers (Lin et al., 2013) introduces the idea of

taking the average of each feature map and feeding

the resulting vector directly into the softmax layer in-

stead of adding fully connected layers on top of the

convolutional neural network. However, in our exper-

iments, the addition of a few more layers after GAP

produced a slight increase in validation accuracy com-

pared to the same setup using max pooling layers. The

VGGNet and Xception were designed to classify im-

ages into 1000 different classes, while our problem

involves only two classes. With few classes, it is pos-

sible to reduce the number of neurons in each fully

connected layer without any decrease in model accu-

racy.

A Dropout (Hinton et al., 2012a) layer was in-

cluded, following each neuron layer, with a ﬁxed

dropout rate of 50%. The only exception was the out-

put layer. A dropout rate of 50% means randomly dis-

abling half of the neuron connections in every train-

ing batch. This approach helps to prevent overﬁtting

and complex co-adaptations on training data (Hinton

et al., 2012b).

The output layer is composed of two neurons that

use softmax activation function. The main property

of softmax is to produce a distribution of probabilities

in the output of the neural network based on neuron

logits. The softmax function is given by equation 1:

softmax(z

) =

∑

j=1

(1)

where z

is the vector formed by the K logits of the

output layer.

Figure 4 shows a summary of the top layers of the

three architectures.

2.4 Image Normalization

The lymphocytes present in the images have a major

axis of 223 ± 43 pixels on average. Due to this at-

tribute, we decided to crop 100 pixels from each bor-

der, reducing the original image size from 450×450

to 250×250. By performing this operation, it was

possible to reduce the image size without losing sub-

stantial cell area and to avoid resizing algorithms. Af-

ter cropping, the pixels were converted to ﬂoat by di-

viding their values by 255.0, and the channel mean

value was subtracted.

2.5 Training

All training was done in a virtual machine from

Google Cloud Platform with an Intel Core i5 2.40

GHz, 20 GB of RAM, and an NVIDIA Tesla T4

graphic card.

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

688

Figure 4: Top layers of the convolutional networks used in

this work: (a) original Xception top layers; (b) Xception

variant used in our work; (c) Original VGG16 and VGG19

top layers, and; (d) VGG16 and VGG19 variants used in our

work.

Along with softmax, the binary cross-entropy loss

was utilized. Cross-entropy loss measures the similar-

ity of a classiﬁcation model that outputs a probability

value between 0 and 1 for each class. This loss pe-

nalizes divergence of the predicted probability q from

its target probability distribution p as deﬁned in Equa-

tion 2. The loss was optimized using Adam with the

same hyperparameter values proposed by Kingma and

Ba (Kingma and Ba, 2014).

H(p, q) = −

∑

log(q

) (2)

We applied a regularization term only to the fully-

connected layers. The regularization term is com-

posed of L

and L

norms. The L

regularization is

the sum of the absolute values of the weight matrix

of the neuron layer, and L

regularization is the sum

of all squared weight values of the same matrix. We

combine these two norms into a single regularization

term, λ, to simplify the ﬁne-tuning process. The fully-

connected network loss can be rewritten as in Equa-

tion 3:

Loss = Cross-Entropy+λ

∑

+ λ

∑

Loss = Cross-Entropy+λ

∑

(3)

where W

is the layer weight matrix with coefﬁcients

i, λ

are L

and L

regularization terms respec-

tively and λ = λ

= λ

. The regularization term λ

was adjusted in order to obtain the lowest validation

loss.

A learning rate schedule was also used, with initial

and ﬁnal learning rates that exponentially decay over

400 epochs. Their values were chosen to make the

training process of each CNN stable.

3 RESULTS AND DISCUSSION

The Test dataset is unbalanced, which can result in

misleading accuracy. To evaluate the performance of

the classiﬁers, the primary metric used was the F1-

score. An advantage of the F1-score is the possi-

bility of comparing our results with those obtained

by the teams that participated in SBILab’s challenge,

which have used the same dataset and also use the

F1-score as the evaluation metric for ranking pur-

poses. The results can be found in a book published

by Gupta (Gupta and Gupta, 2019).

The accuracy, precision, sensitivity (also known

as recall), and speciﬁcity obtained by the best per-

forming convolutional neural networks when applied

to the Test set are shown in Table 2.

In this study, the best model achieved an F1-

score of 92.60%, precision of 91.14%, sensitivity of

94.10%, and speciﬁcity of 90.86% for the malignant

class. This result was obtained with a VGG16 net-

work using as the prediction stage the following se-

quence of layers: one GAP layer, a fully connected

layer with 512 neurons using ReLU, a 50% dropout

layer, another fully connected layer with 512 neurons

using ReLU, another 50% dropout layer and, on the

top, a two-neuron layer using softmax, as shown in

Figure 4(d). The network was trained from scratch.

A learning rate schedule was also used, with an ini-

tial learning rate of 0.00001 exponentially decaying

to 0.000001 over 400 epochs. The regularization term

λ was set to 0.0001. The validation and training accu-

racies are shown in Figure 5.

The ﬁrst runner-up achieved an F1-score of

91.75%, a precision of 90.06%, a sensitivity of

93.51%, and a speciﬁcity of 86.68% for the malignant

class. This result was obtained by a VGG19 network

using the same setup as in the VGG16. The validation

and training accuracy during the training are shown in

Figure 6.

Table 2: Performance of the trained Convolutional Neural

Network architectures in Test set.

Accuracy Precision Sensitivity Speciﬁcity F1-Score

VGG16 92.48% 91.14% 94.10% 90.86% 92.60%

VGG19 91.59% 90.06% 93.51% 86.68% 91.75%

Xception 90.41% 87.64% 94.10% 86.73% 90.76%

Classiﬁcation of Normal versus Leukemic Cells with Data Augmentation and Convolutional Neural Networks

689

Table 3: Performance of participants in C-NMC challenge hosted by SBILab.

Participant in

SBILab challenge

F1-score Methodology

Our model

92.60%

Train a VGG16 architecture from scratch

(Pan et al., 2019)

92.50%

Transfer learning Resnets in a neighborhood-correction algorithm

(Honnalgere and Nayak, 2019)

91.70%

Transfer learning with a VGG16 architecture

(Xiao et al., 2019)

90.30%

Deep multi-model ensemble network (various convolutional neural networks)

(Verma and Singh, 2019)

89.47%

Transfer learning with a MobileNetV2 architecture

(Prellberg and Kramer, 2019)

87.89%

Training from scratch a ResNeXt50 architecture

(Shah et al., 2019)

87.58%

Transfer learning with a combination of convolutional and recurrent neural networks

(Marzahl et al., 2019)

87.46%

Transfer learning with a ResNet18 architecture

(Ding et al., 2019)

86.74%

Training from scratch InceptionV3, DenseNet and InceptionResNetV2 architectures

(Kulhalli et al., 2019)

85.70%

Training from scratch ResNeXt50 and ResNeXt101 architectures

(Liu and Long, 2019)

84.00%

Transfer learning with Inception and ResNets architectures

(Khan and Choo, 2019)

81.79%

Transfer learning with ResNets and SENets

Figure 5: VGG16 training.

Figure 6: VGG19 training.

The second runner-up achieved an F1-score of

90.76%, a precision of 87.64%, a sensitivity of

94.10%, and a speciﬁcity of 86.73% for the malig-

nant class. This result was obtained by an Xception

network using as a prediction stage the following se-

quence of layers: one GAP layer, a fully connected

layer with 2048 neurons using ReLU, a 50% dropout

layer, and, on the top, a two-neuron layer using soft-

max, as shown in Figure 4(b). The network was again

trained from scratch with a learning rate schedule

with an initial learning rate of 0.000005 exponentially

decaying to 0.000001 over 400 epochs. The regular-

ization term λ was set to 0.0007. The validation and

Figure 7: Xception training.

training accuracies are shown in Figure 7.

Compared to the results in Table 3, we achieved a

top score result with our methodology using standard

neural network architectures with simple additions.

The F1-score shown in Table 3 published by the par-

ticipants were obtained by models trained and eval-

uated with images from “C-NMC training data” and

“C-NMC test preliminary phase data”. As mentioned

in Subsection 2.1, we used only “C-NMC training

data” to train and evaluate our models.

We evaluated the three models with the competi-

tion Test data, contained in “C-NMC test ﬁnal phase

data”, by submitting the results to the competition

website. Since this dataset is unlabeled, we were un-

able to compute other performance metrics. The only

metric provided by the competition leaderboard is the

Weighted F1-Score. The results obtained by the three

models in the competition leaderboard are shown in

Table 4. The ﬁnal result of ISBI 2019 challenge is

shown in Table 5.

The highest result, 86.35177019%, obtained by

the Xception model could place the proposed method-

ology in the 8th position on challenge’s ranking.

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

690

Table 4: Performance of CNN architectures in Competition

Leaderboard.

Architecture Weighted F1-Score

Xception 86.35177019%

VGG16 83.24786937%

VGG19 82.59010758%

Table 5: Top entries of C-NMC 2019 Challenge pub-

lished (SBILab, 2019b).

Name Rank Weighted F1-Score

Yongsheng Pan 1 0.910

Ekansh Verma 2 0.894

Jonas Prellberg 3 0.889

Fenrui Xiao 4 0.885

Tian Shi 5 0.879

Ying Liu 6 0.876

Salman Shah 7 0.866

Yifan Ding 8 0.855

Xinpeng Xie 9 0.848

4 CONCLUSIONS AND FUTURE

WORK

Our results indicate that the proposed methodology

based on a convolutional neural network is able to

classify lymphocyte images into malignant or healthy

with high accuracy. Simple modiﬁcations in conven-

tional CNN architectures were enough to create clas-

siﬁers with results similar to complex methodologies.

In the literature, many methods were tested with

only a few sample images or with private datasets. On

the other hand, our study was done with a large and

public dataset. Therefore the result obtained is more

general and can easily be replicated.

An F1-score of 92.60% lacks conﬁdence for dis-

ease diagnosis but can serve as a tool to assist on-

cologists. In conclusion, the proposed method in this

study is a technique with high performance in classiﬁ-

cation of cancerous lymphocyte images which can be

used complementarily in immunophenotyping.

REFERENCES

Ahmed, N., Yigit, A., Isik, Z., and Alpkocak, A. (2019).

Identiﬁcation of leukemia subtypes from microscopic

images using convolutional neural network. Diagnos-

tics, 9(3):104.

Chollet, F. (2017). Xception: Deep learning with depthwise

separable convolutions. In 2017 IEEE Conference

on Computer Vision and Pattern Recognition (CVPR).

IEEE.

Dhillon, A. and Verma, G. K. (2019). Convolutional neu-

ral network: a review of models, methodologies and

applications to object detection. Progress in Artiﬁcial

Intelligence, 9(2):85–112.

DI-UNIMI (2020). ALL-IDB: Acute Lymphoblastic

Leukemia Image Database for Image Processing.

https://homes.di.unimi.it/scotti/all/.

Ding, Y., Yang, Y., and Cui, Y. (2019). Deep learning for

classifying of white blood cancer. In Lecture Notes in

Bioengineering, pages 33–41. Springer Singapore.

Duggal, R., Gupta, A., Gupta, R., and Mallick, P. (2017).

SD-layer: Stain deconvolutional layer for CNNs

in medical microscopic imaging. In Medical Im-

age Computing and Computer Assisted Intervention -

MICCAI, pages 435–443. Springer International Pub-

lishing.

Duggal, R., Gupta, A., Gupta, R., Wadhwa, M., and

Ahuja, C. (2016). Overlapping cell nuclei segmen-

tation in microscopic images using deep belief net-

works. In Proceedings of the Tenth Indian Conference

on Computer Vision, Graphics and Image Processing

- ICVGIP. ACM Press.

Gupta, A. and Gupta, R., editors (2019). ISBI 2019 C-

NMC Challenge: Classiﬁcation in Cancer Cell Imag-

ing. Springer Singapore.

Gupta, R., Mallick, P., Duggal, R., Gupta, A., and Sharma,

O. (2017). Stain color normalization and segmenta-

tion of plasma cells in microscopic images as a pre-

lude to development of computer assisted automated

disease diagnostic tool in multiple myeloma. Clinical

Lymphoma Myeloma and Leukemia, 17(1):e99.

Hinton, G. E., Krizhevsky, A., Sutskever, I., and Srivastva,

N. (2012a). System and method for addressing over-

ﬁtting in a neural network. Patent US9406017B2 as-

signed to Google LLC.

Hinton, G. E., Srivastava, N., Krizhevsky, A., Sutskever, I.,

and Salakhutdinov, R. R. (2012b). Improving neural

networks by preventing co-adaptation of feature de-

tectors.

Hoffbrand, A. V. and Moss, P. A. H. (2013). Essential

Haematology. Wiley, New Jersey, USA, 6 edition.

Honnalgere, A. and Nayak, G. (2019). Classiﬁcation of nor-

mal versus malignant cells in b-ALL white blood can-

cer microscopic images. In Lecture Notes in Bioengi-

neering, pages 1–12. Springer Singapore.

Khan, M. A. and Choo, J. (2019). Classiﬁcation of can-

cer microscopic images via convolutional neural net-

works. In Lecture Notes in Bioengineering, pages

141–147. Springer Singapore.

Kingma, D. P. and Ba, J. (2014). Adam: A method for

stochastic optimization.

Kulhalli, R., Savadikar, C., and Garware, B. (2019). To-

ward automated classiﬁcation of B-acute lymphoblas-

tic leukemia. In Lecture Notes in Bioengineering,

pages 63–72. Springer Singapore.

Labati, R. D., Piuri, V., and Scotti, F. (2011). ALL-IDB:

The Acute Lymphoblastic Leukemia Image Database

for Image Processing. In 18th IEEE International

Classiﬁcation of Normal versus Leukemic Cells with Data Augmentation and Convolutional Neural Networks

691

Conference on Image Processing - ICIP, pages 2045–

2048. IEEE.

Lin, M., Chen, Q., and Yan, S. (2013). Network in network.

Liu, Y. and Long, F. (2019). Acute lymphoblastic leukemia

cells image analysis with deep bagging ensemble

learning. In Lecture Notes in Bioengineering, pages

113–121. Springer Singapore.

Marzahl, C., Aubreville, M., Voigt, J., and Maier, A. (2019).

Classiﬁcation of leukemic b-lymphoblast cells from

blood smear microscopic images with an attention-

based deep learning method and advanced augmenta-

tion techniques. In Lecture Notes in Bioengineering,

pages 13–22. Springer Singapore.

Mishra, S., Majhi, B., and Sa, P. K. (2019). Texture feature

based classiﬁcation on microscopic blood smear for

acute lymphoblastic leukemia detection. Biomedical

Signal Processing and Control, 47:303–311.

Mishra, S., Sharma, L., Majhi, B., and Sa, P. K. (2016).

Microscopic image classiﬁcation using DCT for the

detection of acute lymphoblastic leukemia (ALL).

In Advances in Intelligent Systems and Computing,

pages 171–180. Springer Singapore.

MoradiAmin, M., Memari, A., Samadzadehaghdam, N.,

Kermani, S., and Talebi, A. (2016). Computer aided

detection and classiﬁcation of acute lymphoblastic

leukemia cell subtypes based on microscopic im-

age analysis. Microscopy Research and Technique,

79(10):908–916.

Moshavash, Z., Danyali, H., and Helfroush, M. S. (2018).

An automatic and robust decision support system

for accurate acute leukemia diagnosis from blood

microscopic images. Journal of Digital Imaging,

31(5):702–717.

Mourya, S., Kant, S., Kumar, P., Gupta, A., and Gupta, R.

(2018). LeukoNet: DCT-based CNN architecture for

the classiﬁcation of normal versus Leukemic blasts in

B-ALL Cancer.

Mourya, S., Kant, S., Kumar, P., Gupta, A., and Gupta, R.

(2019). ALL challenge dataset of ISBI.

Pan, Y., Liu, M., Xia, Y., and Shen, D. (2019).

Neighborhood-correction algorithm for classiﬁcation

of normal and malignant cells. In Lecture Notes in

Bioengineering, pages 73–82. Springer Singapore.

Prellberg, J. and Kramer, O. (2019). Acute lymphoblastic

leukemia classiﬁcation from microscopic images us-

ing convolutional neural networks. In Lecture Notes

in Bioengineering, pages 53–61. Springer Singapore.

Putzu, L., Caocci, G., and Ruberto, C. D. (2014). Leuco-

cyte classiﬁcation for leukaemia detection using im-

age processing techniques. Artiﬁcial Intelligence in

Medicine, 62(3):179–191.

Rawat, J., Singh, A., Bhadauria, H. S., Virmani, J., and De-

vgun, J. S. (2017). Classiﬁcation of acute lymphoblas-

tic leukaemia using hybrid hierarchical classiﬁers.

Multimedia Tools and Applications, 76(18):19057–

19085.

Rehman, A., Abbas, N., Saba, T., ur Rahman, S. I.,

Mehmood, Z., and Kolivand, H. (2018). Classiﬁcation

of acute lymphoblastic leukemia using deep learning.

Microscopy Research and Technique, 81(11):1310–

1317.

Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh,

S., Ma, S., Huang, Z., Karpathy, A., Khosla, A.,

Bernstein, M., Berg, A. C., and Fei-Fei, L. (2015).

ImageNet Large Scale Visual Recognition Challenge.

International Journal of Computer Vision (IJCV),

115(3):211–252.

SBILab (2019a). Classiﬁcation of Normal

vs Malignant Cells in B-ALL White

Blood Cancer Microscopic Images: ISBI.

https://competitions.codalab.org/competitions/20395.

SBILab (2019b). Classiﬁcation of Normal vs Malignant

Cells in B-ALL White Blood Cancer Micro-

scopic Images: Top Entries. https://competitions.

codalab.org/competitions/20395#learn the details-

top-entries-of-the- challenge.

SBILab (2020). Signal processing and Bio-medical Imag-

ing Lab. http://sbilab.iiitd.edu.in/.

Shaﬁque, S. and Tehsin, S. (2018). Acute lymphoblastic

leukemia detection and classiﬁcation of its subtypes

using pretrained deep convolutional neural networks.

Technology in Cancer Research & Treatment, 17.

Shah, S., Nawaz, W., Jalil, B., and Khan, H. A. (2019).

Classiﬁcation of normal and leukemic blast cells in

B-ALL cancer using a combination of convolutional

and recurrent neural networks. In Lecture Notes in

Bioengineering, pages 23–31. Springer Singapore.

Simonyan, K. and Zisserman, A. (2014). Very deep convo-

lutional networks for large-scale image recognition.

Verma, E. and Singh, V. (2019). ISBI challenge 2019: Con-

volution neural networks for b-ALL cell classiﬁcation.

In Lecture Notes in Bioengineering, pages 131–139.

Springer Singapore.

Xiao, F., Kuang, R., Ou, Z., and Xiong, B. (2019).

DeepMEN: Multi-model ensemble network for b-

lymphoblast cell classiﬁcation. In Lecture Notes in

Bioengineering, pages 83–93. Springer Singapore.

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

692