SAKARYA UNIVERSITY JOURNAL OF COMPUTER AND INFORMATION SCIENCES

VOL. 5, NO. 3, DECEMBER 2022

DOI: 10.35377/saucis.05.03.1196934
Research/Review Article

Automatic Classification of White Blood Cells Using Pre-Trained

Deep Models

Oguzhan Katar1, Ilhan Firat Kilincer2

1
Corresponding Author; Firat University, Department of Software Engineering; [email protected]; 0000-0002-
5628-3543; +90 424 607 81 07
2
Firat University, Department of Informatics; [email protected]; 0000-0001-8090-4998

Received 31 October 2022; Revised 1 November 2022; Accepted 22 December 2022; Published online 31 December 2022

Abstract

White blood cells (WBCs), which are a crucial component of the immune system, help our body defend against
infections and other diseases. Some diseases may cause our body to produce fewer WBCs than it requires.
Therefore, WBCs are of great importance in medical imaging. Artificial intelligence-based computer systems can
assist experts in analyzing WBCs. In this study, we proposed an approach for the automatic classification of WBCs
into five different classes using a pre-trained model. We trained ResNet-50, VGG-19, and MobileNet-V3-Small
pre-trained models with ImageNet weights. For the training, validation, and testing processes of the models, we
used a public dataset containing 16,633 images with an uneven class distribution. While the ResNet-50 model
achieved an accuracy of 98.79%, the VGG-19 model achieved an accuracy of 98.19%, and the MobileNet-V3-
Small model achieved the highest accuracy rate at 98.86%. When examining the predictions of the MobileNet-
V3-Small model, we observed that it was not affected by class dominance and was able to correctly classify even
the least sampled class images in the dataset. In addition to the high accuracy achieved in the classification of
WBCs using the proposed pre-trained deep learning models, we also applied the Grad-CAM method to further
understand and interpret the model's predictions.

Keywords: white blood cells, classification, pre-trained models, artificial intelligence, Grad-CAM

1. Introduction

Blood is a vital fluid that helps to nourish the body, maintain acid-base balance, transport hormones,
and maintain salt and water balance. Blood consists of three types of cells: erythrocytes, platelets, and
leukocytes [1].
Erythrocytes, the most abundant type of blood cell, contain a substance called hemoglobin, which is
responsible for transporting oxygen in the body [2]. Oxygen, inhaled into the lungs through respiration
and then entering the blood, can be transported to all body tissues with the help of hemoglobin in
erythrocytes. Adequate oxygen access to each cell in the body depends on the sufficient number and
function of erythrocytes in the blood. Erythrocytes, which are reddish in color and therefore also referred
to as red blood cells, obtain their color from the iron mineral in the structure of hemoglobin [3].
Platelets are cell fragments that are formed by the disintegration of cells called megakaryocytes in the
bone marrow tissue located in the center of our bones after they mature and enter the blood [4]. Platelets
play a vital role in regulating certain chemical reactions that occur in the blood due to the biochemical
substances they contain [5]. However, their primary function is in the case of bleeding due to injury to
blood vessels; they help to quickly close and repair the wounded area.
Leukocytes, also known as white blood cells (WBCs), are an important part of the immune system and
a group of cells that protect the body against infections [6]. When the body encounters foreign
organisms, they reproduce rapidly. The primary function of leukocytes is to identify and eliminate
antigens such as bacteria, viruses, fungi, and poisonous toxins that have entered the body in various
ways. Leukocytes consist of five different types of WBCs, each with its own specific functions:
● Basophils, which are the least common type of leukocyte in the body, fight infections and
parasitic infections. By releasing histamine during allergic reactions, basophils enable the body
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

to produce an antibody called immunoglobulin E. Additionally, by secreting heparin, they

increase the fluidity of the blood [7].
● Eosinophils produce enzymes that destroy parasites that cause inflammatory and allergic
reactions in the body [8].
● Monocytes are produced in the bone marrow and then enter the bloodstream. These cells are
called monocytes when in the bloodstream, but within a few hours, they leave the circulatory
system and enter the tissues. The monocyte cells that reach the tissue are called macrophages.
They eliminate microorganisms that cause infections and clean up dead cells [9].
● Lymphocyte cells, which are produced in the bone marrow and lymph tissue, secrete chemicals
called lymphokines against foreign organisms in the body, stimulating other immune system
cells and allowing them to attack the foreign organism [10].
● Neutrophils are the first precursor cells to reach foreign organisms that cause infections in the
body. They release and digest chemical enzymes to combat foreign organisms [11].
Leukemia, anemia, cancer, and various other diseases can be diagnosed through the analysis of WBCs
[12]. This analysis is often conducted using a peripheral blood smear, which is a common laboratory
method. To obtain a sample, a healthcare provider draws blood from a patient's finger or toe using a
sterile needle, and the sample is then examined in a laboratory to create a peripheral blood film [13].
This film is manually analyzed by a specialist to identify signs of disease. However, manual analysis
can be time-consuming and laborious for experts. As a result, computer-aided systems have been
developed to assist with the classification of WBCs. With the advancement of hardware technology, the
use of artificial intelligence (AI) in this field has increased. AI-based systems, also known as decision
support systems, are designed to minimize errors caused by human factors and are used in various
sectors, including healthcare. For example, decision support systems have been successfully used to
detect COVID-19 through chest computed tomography images and to detect brain tumors through brain
magnetic resonance imaging without human intervention [14].
Many studies have been carried out for the automatic classification of WBCs by AI-based systems. In
the study [15], researchers proposed a system that uses the DenseNet-121 model to classify different
types of WBCs. A publicly available dataset including eosinophil, lymphocyte, monocyte, and
neutrophil classes was used for model training. The dataset contains 12,444 different samples with a
resolution of 320×240px. The normalization process was applied to the dataset samples to speed up
model training. The number of dataset samples has been increased with data augmentation techniques
such as flipping, rotation, brightness, and zooming. The dataset samples were resized to a resolution of
224×224px. After the pre-processing steps, 20,050 WBCs images were obtained, including synthetic
images. The model is trained for 10 epochs with the help of the Adam optimizer. Four different training
processes were performed and the batch size value was changed to 8, 16, 32, and 64 in each training.
The model, which was trained with 8 batch sizes, achieved 98.84% accuracy, 99.33% precision, 98.85%
sensitivity, and 99.61% specificity values during the test phase, and achieved more successful results
compared to other models.
In the study [16], researchers proposed an approach that can classify WBCs from microscopic blood
images. The researchers used a publicly available dataset of images with different values in resolutions
ranging from 350×236px to 2592×1944px. AlexNet, ResNet-101, and GoogleNet models were trained
to detect five different classes: basophil, eosinophil, lymphocyte, monocyte, and neutrophil. While the
dataset samples are resized to 227×227px resolution for training the AlexNet model, this value is
224×224px for the training of the GoogleNet and ResNet-101 models. To compare the success of the
pre-trained models in classifying WBCs, 178 test images were given to the relevant models as input.
The AlexNet model achieved better results compared to other models with 96.63% accuracy, 97.85%
specificity, and 89.18% sensitivity rates.
In one study [17], researchers designed a deep convolutional neural network (CNN) model to classify
microscopic images of WBCs. They proposed a new data augmentation method based on feature
concentration to enhance the dataset and address the small number of samples. The training, validation,

463
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

and testing processes for the CNN model, which was designed to automatically classify the neutrophil,
lymphocyte, monocyte, eosinophil, and basophil classes, were carried out using a special dataset
provided by Sichuan Meisheng Biotech Company. This dataset consists of 8600 leukocyte images with
a resolution of 1024×768px collected from various individuals. These images were divided into
217×217px pieces, resulting in a total of 11,658 sub-images. 80% of the dataset samples were reserved
for training, while the remaining 20% were used for validation. The proposed model achieved an average
test accuracy of 97.6% in classifying the five different WBCs.
In another study [18], researchers proposed an approach for classifying WBCs in microscopic images.
Samples from a publicly available dataset containing a total of 352 images were augmented using
various image augmentation techniques, resulting in 12,444 images. The dataset included samples
belonging to the eosinophil, lymphocyte, monocyte, and neutrophil classes. A seven-layer convolutional
neural network with an input size of 120×160px was created to automatically classify these samples. To
this end, all of the dataset samples were resized to 120×160px. The proposed model was subjected to
two different training processes to examine its binary and multiclass classification performance. In
binary classification, a mononuclear class was created using eosinophil and neutrophil samples, and a
polynuclear class was created using lymphocyte and monocyte samples. The model achieved an
accuracy of 96.30% in binary classification and 87.93% accuracy in multiclass classification.
In another study [19], the researchers proposed a system that can simultaneously detect and classify
WBCs in an image. This system is based on the F-RCNN and YOLOv4 architectures. The models were
trained on samples from the Blood Cell Count Dataset (BCCD), which includes samples of four different
WBCs: neutrophils, eosinophils, monocytes, and lymphocytes. The F-RCNN model achieved an
accuracy of 96.25% and the YOLOv4 model achieved 95.75% accuracy during the testing phase.
In yet another study [20], the researchers proposed a U-Net-based approach for WBCs segmentation. In
the U-Net encoder network, ResNet-50 blocks were integrated instead of the default layers, and squeeze-
and-excitation blocks were added to the decoder network. The training and testing stages of the model
were conducted using samples from the BCISC and LISC datasets. Using various data augmentation
techniques, the number of samples for each dataset was increased to 10,000. The dataset samples were
divided into 80% for training, 10% for validation, and 10% for testing. The ResNet-50-based U-Net
model was trained for 200 epochs with a batch size of 8 and Adam optimization. It was reported that
the model achieved a Dice score of 98.13% and a mean Intersection over Union (mIoU) rate of 96.36%
during the testing phase using the BISC dataset samples.
The primary objective of this study is to use deep learning to automatically detect WBCs from
microscopic blood images, thereby assisting specialists in the early diagnosis of diseases related to
WBCs counts. The main contributions of this study are as follows:
● Demonstrating the effectiveness of existing deep learning models on a new dataset.
● Achieving high performance on a non-uniformly distributed dataset without using data
augmentation for WBCs classification.
● Visualizing, using Gradient-weighted Class Activation Mapping (Grad-CAM), which pixel
areas the deep learning models focus on during the decision-making phase, thereby providing
an explainable structure for pre-trained models.
● Reducing human errors and subjectivity by using deep learning structures to perform these
tasks, which are currently carried out by experts visually.
The remainder of this paper is organized as follows: Section 2 presents the proposed method for this
study, including the dataset used, pre-trained deep learning models, classification performance
measures, and the Grad-CAM algorithm. Section 3 presents the parameters and environments used in
the training phase, the numerical values of the model during the training phase, the test phase
predictions, and performance values. The discussion and conclusion sections of the study are presented
in Section 4 and Section 5, respectively.

2. Material and Methods

464
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

An approach has been proposed for the deep learning-based automated classification of WBCs from
microscopic blood images. The block representation of the proposed method is given in Figure 1.

Basophil
Basophil

Image Cropping

Image Resizing
Monocyte
Eosinophil

Lymphocyte Eosinophil
Pre-trained Model Lymphocyte

Neut rophil
Monocyte

Neutrophil

Microscopic Blood Image Dataset

Grad-CAM

Output

Figure 1 A Block Representation of The Proposed Method

In the proposed method, the image given as input to the deep learning model is classified as a basophil,
eosinophil, lymphocyte, monocyte, or neutrophil at the model's output. The choice of dataset and model
is critical for achieving high success rates in this classification process. The quality of the dataset directly
affects the performance of the deep learning model, and therefore it is important that it is created or
verified by experts. This can be a resource and time-intensive process. However, several researchers
have created and publicly shared WBCs datasets, as listed in Table 1.
Table 1 Publicly Available WBCs Datasets
Dataset Basophil Eosinophil Lymphocyte Monocyte Neutrophil Total
LISC [21] 54 42 59 55 56 266
BCCD [22] 3 86 33 19 208 349
MISP [23] 0 42 36 33 38 149
ALL-IDB [24] 1 2 60 3 18 84
Zheng et al. [25] 1 22 53 48 176 300
Raabin-WBC [26] 301 1066 3609 795 10,862 16,633

2.1 Dataset
In this study, the Raabin-WBC dataset [26] was used for the training, validation, and testing of the
models. The Raabin-WBC dataset was created using 72 peripheral blood films collected from Shariati
Hospital, which were examined using Olympus Cx18 and Zeiss microscopes. A total of 16,633 WBCs
images with a resolution of 575×575px were obtained, and these images were labeled by two experts:
301 were labeled as basophils (Bas), 1066 as eosinophils (Eos), 3609 as lymphocytes (Lym), 795 as
monocytes (Mon), and 10,862 as neutrophils (Neu). Samples of each class in the dataset are shown in
Figure 2.

(a) Basophil (b) Eosinophil (c) Lymphocyte (d) Monocyte (e) Neutrophil

Figure 2 Dataset Samples [26]

Upon examination of the class-based distribution of samples in the Raabin-WBC dataset, it was
observed that the Neu class is dominant. Data augmentation methods, which involve creating synthetic
images, can be used to balance the distribution of classes. However, in this study, no data augmentation
was performed in order to test the performance of pre-trained models under challenging conditions.

465
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

2.2 Pre-trained Models

Transfer learning is a machine learning technique that involves using the weights of a previously trained
model as initial weights in the training phase of CNN. This allows the model, which was previously
trained on a task, to be reused for different tasks. Transfer learning is highly effective for achieving good
performance with a small amount of data. It is now a widely used method, especially for tasks related
to image or natural language processing, as it allows researchers to use pre-trained models that have
already learned how to classify images and have learned general features such as edges and shapes.
Examples of pre-trained models that are often used as the basis for transfer learning include ResNet
[27], VGG [28], and MobileNet [29], which were trained using the ImageNet [30] database. Pre-trained
models can be grouped into three categories based on the number of parameters they contain: low (less
than 15 M), medium (between 15 M - 70 M), and high (more than 70 M). Information on the pre-trained
models is provided in Table 2 [31].
Table 2 Pre-trained Models Used in This Study [31]
Model Default Input Size Parameters (Million) Category
ConvNeXtXLarge 224×224 350.1 High
ConvNeXtLarge 224×224 197.7 High
VGG-19 224×224 143.7 High
VGG-16 224×224 138.4 High
EfficientNetV2L 480×480 119 High
NASNetLarge 331×331 88.9 High
ConvNeXtBase 224×224 88.5 High
EfficientNetB7 600×600 66.7 Medium
ResNet152 224×224 60.4 Medium
ResNet152V2 224×224 60.4 Medium
InceptionResNetV2 299×299 55.9 Medium
EfficientNetV2M 480×480 54.4 Medium
ConvNeXtSmall 224×224 50.2 Medium
ResNet101 224×224 44.7 Medium
ResNet101V2 224×224 44.7 Medium
EfficientNetB6 528×528 43.3 Medium
EfficientNetB5 456×456 30.6 Medium
ConvNeXtTiny 224×224 28.6 Medium
ResNet50 224×224 25.6 Medium
ResNet50V2 224×224 25.6 Medium
InceptionV3 299×299 23.9 Medium
Xception 299×299 22.9 Medium
EfficientNetV2S 384×384 21.6 Medium
DenseNet201 224×224 20.2 Medium
EfficientNetB4 380×380 19.5 Medium
EfficientNetV2B3 300×300 14.5 Low
DenseNet169 224×224 14.3 Low
EfficientNetB3 300×300 12.3 Low
EfficientNetV2B2 260×260 10.2 Low
EfficientNetB2 260×260 9.2 Low
EfficientNetV2B1 240×240 8.2 Low
DenseNet121 224×224 8.1 Low
EfficientNetB1 240×240 7.9 Low
EfficientNetV2B0 224×224 7.2 Low
Mobilenet_v3_large 224×224 5.4 Low
NASNetMobile 224×224 5.3 Low
EfficientNetB0 224×224 5.3 Low
MobileNet 224×224 4.3 Low
MobileNetV2 224×224 3.5 Low
Mobilenet_v3_small 224×224 2.9 Low

466
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

To directly assess the effect of the number of parameters on model performance, three pre-trained
models were randomly selected from the categories specifically created for this study: ResNet-50, VGG-
19, and MobileNet-V3-Small.

2.3 Performance Metrics

Various metrics can be calculated using True Positive (TP), False Positive (FP), True Negative (TN),
and False Negative (FN) to evaluate the performance of models. In this study, four metrics were used
to evaluate the models for each class. These metrics and their corresponding equations are as follows:
● Accuracy is a performance metric that measures the percentage of correct predictions made by
a classification model. It is the most widely used performance metric, but it may not fully reflect
the performance of a model and can sometimes be misleading. For example, in a dataset where
some classes are more represented than others, accuracy may not be a sufficient metric.
● Precision measures the percentage of predictions made by a model that are correct. The main
difference between precision and accuracy is that precision only considers correct predictions,
while accuracy considers all predictions. Therefore, precision is often a more precise metric and
is given greater consideration when evaluating the performance of classification models.
● Sensitivity is a performance metric that measures the success of a classification model. It shows
the percentage of data that the model predicted correctly. The main difference with other metrics
is that sensitivity only evaluates correct predictions. For example, a model may have low
sensitivity even though it has high accuracy. In this case, most of the data that the model predicts
correctly are misclassified data, indicating that the model is not performing well.
● The F-1 score is a combination of sensitivity and precision ratios, used to evaluate the
performance of a classification model, especially for multi-label data. The advantage of the F-
1 score is that it does not rely solely on accuracy values, allowing it to show whether the model
has balanced performance for all classes.
𝑇𝑃 (1)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 (𝑃) =
𝑇𝑃 + 𝐹𝑃
𝑇𝑃 (2)
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 (𝑆) =
𝑇𝑃 + 𝐹𝑁
𝑇𝑃 + 𝑇𝑁 (3)
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 (𝐴𝑐𝑐) =
𝑇𝑃 + 𝐹𝑃 + 𝑇𝑁 + 𝐹𝑁
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦
𝐹1 𝑆𝑐𝑜𝑟𝑒 (𝐹1) = 2 × (4)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦

2.4 Grad-CAM Algorithm

Grad-CAM is a technique that visualizes the regions of an image that are most important for a CNN
to make a prediction. It allows us to understand which parts of an image a CNN is using to make a
decision, and can be used to generate heatmaps that highlight these regions [32]. Grad-CAM works
by using the gradients of the output of the CNN with respect to the input image to produce a weighted
sum of the feature maps in the final convolutional of the network. The resulting heatmap is then
overlaid on the input image to show which regions had the greatest influence on CNN's prediction.
The architecture of the Grad-CAM algorithm is depicted in Figure 3.

467
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

Figure 3 The architecture of the Grad-CAM [32]

The process for creating a Grad-CAM visualization for a pre-trained CNN is as follows:
1. Feed the input image through the CNN to generate a prediction.
2. Compute the gradients of the output of the CNN with respect to the feature maps in the final
convolutional layer.
3. Take the weighted sum of the feature maps, using the gradients as weights.
4. Resize the resulting heatmap to the size of the input image.
5. Overlay the heatmap on the input image to highlight the regions that were most important
for CNN's prediction.
Grad-CAM is relatively simple to implement and can be used with any CNN, regardless of its
architecture. It is also an efficient method, as it only requires a single forward and backward pass
through the network to generate the visualization. However, there are some limitations to Grad-
CAM. For example, it can only provide visualizations for a single class at a time, and it is sensitive
to the specific layer chosen for visualization. Additionally, the visualizations produced by Grad-
CAM may not always align perfectly with human intuition, as they are based on the internal
representation of the CNN rather than the visual features that a human might use to classify the
image.

3. Experimental Results

The results of models trained to classify WBCs from microscopic blood images are presented in this
section. In addition, an analysis of the experimental findings with performance metrics is shown in the
following sections.

3.1 Experimental Setups

The default input size of the ResNet-50, VGG-19, and MobileNet-V3-Small models used in this study
is 224×224px, so all of the dataset samples were resized to this value. Before training the model, 70%
of the resized dataset samples were randomly divided for use in the training, 20% for validation, and
10% for testing. The visual representation of these processes is shown in Figure 4.

468
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

575px
224px
Basophil

224px
575px

211/60/30

224px
Eosinophil

224px
575px

746/213/107

Image Resizing
224px
Lymphocyte

224px
575px

2526/722/361

224px
Monocyte

224px
575px

556/159/80

224px
Neutrophil
575px

224px

7604/2172/1086

575px

= Train (70%) = Validation (20%) = Test (10%)

Figure 4 Image Resizing and Splitting Method
The pre-trained ResNet-50, VGG-19, and MobileNet-V3-Small models were included in the training
using the Keras library. Since the models will only make predictions for five different classes, the Dense
layers were revised and the softmax activation function was used. The models were compiled with an
Adam optimizer and a learning rate of 0.0001. ImageNet weights were used instead of random initial
weights for the training of the models. The models were trained with a constant batch size of 64, and
training was carried out for a maximum of 50 epochs using the early stopping function. If the monitored
validation accuracy value does not improve for five consecutive epochs, the early stopping function
terminates the training phase and the weights of the epoch with the highest validation accuracy value
are recorded in the '.h5' format. All of these processes were performed in the Google Colab environment.

3.2 Results
Three different deep-learning models were trained with the same parameters. The time required to
complete the model training processes is directly proportional to the number of parameters and layers
they have. The training stages of the models were carried out using the early stopping function, and the
weights of the epoch that achieved the highest validation accuracy were recorded. ResNet-50 reached
the highest validation accuracy after 16 epochs, VGG-19 reached the highest validation accuracy after
23 epochs, and MobileNet-V3-Small reached the highest validation accuracy after 24 epochs. The loss
and accuracy graphs for the models during the training and validation phases are shown in Figure 5.

469
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

0.7 1

0.98
0.6 Train Loss
0.96
Validation Loss
0.5 0.94

Accuracy Value
Loss Value

0.92
0.4

0.9
0.3
0.88

0.2 0.86

0.84 Train Accuracy

0.1
0.82 Validation Accuracy

0 0.8
0 2 4 6 8 10 12 14 16 0 2 4 6 8 10 12 14 16
Epoch Epoch

(a) ResNet-50
1.4 1

1.2 Train Loss

0.9
Validation Loss
1

Accuracy Value
0.8
Loss Value

0.8
0.7
0.6

0.6
0.4

Train Accuracy
0.5
0.2
Validation Accuracy

0 0.4
0 2 4 6 8 10 12 14 16 18 0 5 10 15 20 25
Epoch Epoch

(b) VGG-19
0.6 1

Train Loss 0.98

0.5
Validation Loss 0.96
Accuracy Value

0.4
0.94
Loss Value

0.3 0.92

0.9
0.2

0.88
Train Accuracy
0.1
0.86 Validation Accuracy

0 0.84
0 5 10 15 20 25 0 5 10 15 20 25
Epoch Epoch

(c) MobileNet-V3-Small
Figure 5 Loss and Accuracy Graphs
When the validation accuracy values were examined for the three models that completed training, it was
observed that a rate of more than 95% could be achieved in less than 25 epochs. This is due in part to
the fact that the models were trained using ImageNet weights instead of starting with random weights.
Even though the models were trained with a dataset that is not evenly balanced, the lack of overfitting
indicates the success of the pre-trained models. Performance metrics were used to compare the
classification performance of the three different models trained to classify five different WBCs from
microscopic blood images. For this, images that were not included in the training and validation phases
but were reserved solely for use in the testing phase were given as input to each model. The confusion
matrices generated by the predictions of the models for these inputs are shown in Figure 6.

470
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

Actual Class

Actual Class
Predicted Class Predicted Class

(a) ResNet-50 (b) VGG-19

Actual Class

Predicted Class

(c) MobileNet-V3-Small

Figure 6 Confusion Matrices for Each Model

To evaluate the Grad-CAM outputs of the deep learning model, it is necessary to first assess the
performance of the model during the training phase. This helps to understand the accuracy of the model's
predictions and assess the reliability of the model. When the predictions are analyzed, it is apparent that
the models have learned to classify WBCs. In Table 3, the performance metric values achieved by the
relevant models during the testing phase are provided.
Table 3 The Results of The Pre-trained Models
Basophil Eosinophil Lymphocyte Monocyte Neutrophil
Model
P S F1 P S F1 P S F1 P S F1 P S F1
(%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%) (%)
ResNet-50 100 100 100 100 96.26 98.09 97.78 97.78 97.78 91.46 93.75 92.59 99.54 99.72 99.62

VGG-19 96.55 93.33 94.91 97.19 97.19 97.19 96.95 96.95 96.95 86.51 96.25 91.12 99.72 98.98 99.34
MobileNet-
100 100 100 100 93.45 96.61 98.61 98.33 98.46 92.85 97.50 95.11 99.26 99.63 99.44
V3-Small

Following the training phase, the model should be evaluated using the test data. The resulting outputs
should be carefully analyzed to interpret how the Grad-CAM outputs describe the images and identify
the features that the model considers important. Figure 7 presents the Grad-CAM outputs for a selection
of randomly chosen images from the test set.

471
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

ResNet-50 VGG-19 MobileNet-V3-Small

Input Image Grad-CAM Grad-CAM + Input Grad-CAM Grad-CAM + Input Grad-CAM Grad-CAM + Input

Figure 7 The Grad-CAM outputs

Upon examination of the Grad-CAM outputs, it was observed that all three models effectively identified
the relevant features in the images with a high degree of accuracy. As the model performance is
consistent with expectations, it is not necessary to further adjust the hyperparameters or layer
configurations of the models.

4. Discussion

The detection of WBCs using microscopic blood images is a topic of active research. Table 4 presents
a selection of studies on this subject that have been curated by hand. Yildirim and Cinar [33] employed
AlexNet, ResNet-50, DenseNet-201, and GoogleNet architectures on a dataset comprising 9,663
images. For each model, three different training stages were conducted using original data, data filtered
with a Gaussian filter, and data filtered with a median filter. The highest accuracy rate of 83.44% was
achieved by the DenseNet-201 model trained with Gaussian-filtered data. Ekiz et al. [34] classified
12,442 WBCs images using both a CNN model and a Con-SVM model, with the Con-SVM model found
to be more accurate, achieving an accuracy rate of 85.96%, compared to the CNN model's accuracy rate
of 83.91%. Sharma et al. [15] implemented a deep learning model based on the DenseNet121
architecture for the classification of various types of WBCs. The model was optimized with
normalization and data augmentation and achieved an accuracy of 98.84%. Girdhar et al. [35] proposed
a method that demonstrated the ability to accurately classify WBCs types in a shorter number of
epochs/time compared to other approaches. The performance of the proposed method was evaluated
using the Kaggle dataset, resulting in an overall accuracy of 98.55%. Nahzat et al. [36] aimed to develop
a CNN-based model for the classification of WBCs. They used images of WBCs from the Kaggle dataset
to train and evaluate their proposed model, testing it with various optimizers to determine the best
performance. They also compared the performance of their model with four pre-trained CNN models

472
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

(MobileNetV2, DenseNet121, InceptionV3, and ResNet50) and found that the proposed model, despite
having the lowest number of trainable parameters and training time, outperformed the others with an
accuracy of 99.5%. Karakuş and Özbay [37] used CNN models and combined them with three different
machine learning classifiers. They applied contrast-limited adaptive histogram equalization (CLAHE)
and Gaussian filters to images from the Kaggle dataset, which were then reclassified using the three
CNN networks. The results showed that the classification performance was higher when the images
were preprocessed with these filters compared to the original data. Jung et al. [38] proposed W-Net, a
CNN-based method for the classification of WBCs. To evaluate W-Net, they used a large-scale dataset
of 6,562 real images of the five WBCs types, obtained from The Catholic University of Korea. The
results showed that W-Net achieved an average accuracy of 97%. Wang et al. [39] proposed a deep
CNN called WBC-AMNet for automatically classifying WBCs subtypes based on a focused attention
mechanism. This method uses feature fusion strategies, combining Squeeze-and-Excitation and Gather-
Excite modules, to obtain more localized attention from the CNN. The WBC-AMNet achieved an
overall accuracy of 98.39. They also used Grad-CAM to visualize the attention heatmaps of different
feature maps. Roy and Ameer [40] applied a semantic segmentation technique using a deep learning
network to accurately segment WBCs from microscopic blood images. The proposed model employed
the DeepLabv3+ architecture with a ResNet-50 network as the feature extractor. The model was
evaluated on three different public datasets containing five categories of WBCs, using 10-fold cross-
validation to assess its effectiveness. The average segmentation accuracy achieved by the proposed
model was 96.1% IoU. Wu et al. [20] proposed a WBC image segmentation network based on U-Net
that combines residual networks. The encoder structure of the network uses ResNet50 residual blocks
as the main unit. The proposed model achieved 96.36% mIoU.
Table 4 Comparison of Our Work With Some State-of-the-art Studies
Number of Number of
Study Model Explainability Task Performance
Class Images
Yildirim and 4 (Eos, Lym,
9,663 DenseNet-201 Black-box Classification Acc=83.44%
Cinar [33] Mon, Neu)

Ekiz et al. 4 (Eos, Lym,

12,442 Con-SVM Black-box Classification Acc=85.96%
[34] Mon, Neu)

Sharma et al. 4 (Eos, Lym,

12,444 DenseNet-121 Black-box Classification Acc=98.84%
[15] Mon, Neu)

Girdhar et al. 4 (Eos, Lym,

12,444 CNN Black-box Classification Acc=98.55%
[35] Mon, Neu)

Nahzat et al. 4 (Eos, Lym,

12,444 Hybrid CNN Black-box Classification Acc=99.50%
[36] Mon, Neu)

Karakuş and 4 (Eos, Lym,

12,444 CNN Black-box Classification Acc=97.10%
Özbay [37] Mon, Neu)
5 (Bas, Eos,
Jung et al.
Lym, Mon, 6,562 W-Net Black-box Classification Acc=97.00%
[38]
Neu)
Wang et al. 4 (Eos, Lym,
16,873 WBC-AMNet Grad-CAM Classification Acc=98.39%
[39] Mon, Neu)
5 (Bas, Eos,
Roy and
Lym, Mon, 642 DeepLabv3+ Black-box Segmentation mIoU=96.10%
Ameer [40]
Neu)
5 (Bas, Eos,
Wu et al.
Lym, Mon, 516 U-Net Black-box Segmentation mIoU=96.36%
[20]
Neu)
The 5 (Bas, Eos,
MobileNet-
proposed our Lym, Mon, 16,633 Grad-CAM Classification Acc=98.86%
V3-Small
study Neu)

473
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

In this study, we employed a pre-trained MobileNet-V3-Small model for automated WBCs

classification. Our results demonstrated a high accuracy of 98.86%, which is higher than the accuracy
reported in most other studies. This suggests that the model in our study was able to accurately classify
the images into the appropriate categories. Our study included a larger number of classes (5) compared
to many other studies (which often have only 4 classes). This increased complexity made the task more
challenging and required a more sophisticated model. Our dataset was also relatively large, with 16,663
images, which may have contributed to the robustness and generalizability of our model. We also
employed Grad-CAM as an explainability method to provide insights into the model's decision-making
process and identify any potential biases or weaknesses.
It is worth noting that some other studies have focused on image segmentation, a task distinct from
classification. Image segmentation involves predicting a pixel-level mask for each class in the image,
while classification simply involves predicting a single class label for the entire image. In this study, we
employed the MobileNet-V3-Small model architecture, which may not be optimal for all tasks and
datasets. Alternative model architectures may yield better performance in certain cases. Some other
studies have utilized models with more layers and a greater number of parameters (e.g. DenseNet-201,
DenseNet-121), which may improve performance but also require more computational resources and
may be more prone to overfitting.
The limitations of this study are as follows:
● The dataset consists of only 16,633 images, which may not be sufficient to fully capture the
variability and complexity of the WBCs being analyzed.
● Our study only evaluated the performance of three pre-trained models (ResNet-50, VGG-19,
and MobileNet-V3-Small) on the WBCs classification task.
● The durability of models against changes due to variations in lighting, background, or other
factors that may affect the appearance of WBCs in images has not been validated.
● As k-fold cross-validation was not employed, the model was only evaluated on a single split of
the data.
In future research, it would be beneficial to augment the dataset with a larger number of images that
have a more balanced distribution of classes. This would likely lead to more robust and accurate
classifications. It would also be useful to evaluate the model on a range of different datasets to assess
its generalizability and performance on diverse types of images. While the models in this study
demonstrated high accuracy rates, there is always a potential for further improvement. Additional
research could be conducted to optimize the models and enhance their performance. While the models
in this study demonstrated high accuracy in classifying WBCs, it would be valuable to assess their
performance in real-world settings. This might involve testing the models on images from actual medical
cases or incorporating the models into existing medical imaging systems for use by healthcare
professionals.

5. Conclusion

In recent years, advances in hardware technology have enabled the use of machine learning techniques
in the field of healthcare, specifically in the automatic classification of WBCs using microscopic blood
images. Accurate identification of WBCs is crucial for medical diagnosis and research. This study
proposes a deep learning-based approach for the automatic classification of WBCs using microscopic
blood images and investigates its effectiveness through experiments on a dataset of 16,633 different
WBCs images. Several popular pre-trained models, including MobileNet-V3-Small, were employed for
the deep learning models. The MobileNet-V3-Small model achieved the highest accuracy rate of
98.86%. To understand how the model was making its predictions, we employed a visualization
technique called Grad-CAM to identify the pixel areas that the model was focusing on. The findings of
this study suggest that deep learning may be a useful tool for the automated identification of WBCs in
medical diagnosis and research. However, further research is needed to fully evaluate the robustness

474
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

and generalizability of these results, as well as to explore the potential for using deep learning in other
aspects of medical diagnosis and treatment.

References

[1] C. J. Walsh and C. A. Luer, "Elasmobranch hematology: identification of cell types and practical
applications," The Elasmobranch Husbandry Manual: Captive Care of Sharks, Rays and their
Relatives, pp. 307-323, 2004.
[2] A. Glenn and C. E. Armstrong, "Physiology of red and white blood cells," Anaesthesia & Intensive
Care Medicine, vol. 20, no. 3, pp. 180-174, 2019.
[3] R. Van Zwieten, A. J. Verhoeven and D. Roos, "Inborn defects in the antioxidant systems of
human red blood cells," Free Radical Biology and Medicine, vol. 67, pp. 377-386, 2014.
[4] I. Andia and N. Maffulli, "Platelet-rich plasma for managing pain and inflammation in
osteoarthritis," Nature Reviews Rheumatology, vol. 9, no. 12, pp. 721-730, 2013.
[5] B. Olas and B. Wachowicz, "Role of reactive nitrogen species in blood platelet functions,"
Platelets, vol. 18, no. 8, pp. 555-565, 2007.
[6] M. Habibzadeh, M. Jannesari, Z. Rezaei, H. Baharvand and M. Totonchi, "Automatic white blood
cell classification using pre-trained deep learning models: Resnet and inception," Tenth
international conference on machine vision, vol. 10696, pp. 274-281, 2018.
[7] A. L. Gillen and J. Conrad, "Our Impressive Immune System: More Than a Defense", Faculty
Publications and Presentations, 135, 2014.
[8] H. Kita and B. S. Bochner, "Biology of eosinophils", Middleton’s allergy principles and practice,
vol. 8 pp. 265-279, 2013.
[9] F. Ginhoux and S. Jung, "Monocytes and macrophages: developmental pathways and tissue
homeostasis", Nature Reviews Immunology, vol. 14, no. 6, pp. 392-404, 2014.
[10] Y. Ueda, M. Kondo and G. Kelsoe, "Inflammation and the reciprocal production of granulocytes
and lymphocytes in bone marrow", The Journal of experimental medicine, vol. 201, no. 11, pp.
1771-1780, 2005.
[11] E. Bronze-da-Rocha and A. Santos-Silva, "Neutrophil elastase inhibitors and chronic kidney
disease", International journal of biological sciences, vol. 14, no.10, pp. 1343-1360, 2018.
[12] A. Shahzad, M. Raza, J. H. Shah, M. Sharif and R. S. Nayak, "Categorizing white blood cells by
utilizing deep features of proposed 4B-AdditionNet-based CNN network with ant colony
optimization," Complex & Intelligent Systems, vol. 8, no. 4, pp. 3143-3159, 2022.
[13] L. B. Maedel and K. Doig, "Examination of the peripheral blood film and correlation with the
complete blood count," Hematology: clinical principles and applications, pp. 192-209, 2013.
[14] O. Katar and E. Duman, "Deep Learning Based Covid-19 Detection With A Novel CT Images
Dataset: EFSCH-19," Avrupa Bilim ve Teknoloji Dergisi, vol. 29, pp. 150-155, 2021.
[15] S. Sharma, S. Gupta, D. Gupta, S. Juneja, P. Gupta, G. Dhiman and S. Kautish, "Deep learning
model for the automatic classification of white blood cells," Computational Intelligence and
Neuroscience, 2022.
[16] M. J. Macawile, V. V. Quiñones, A. Ballado, J. D. Cruz and M. V. Caya, "White blood cell
classification and counting using convolutional neural network," 3rd International conference on
control and robotics engineering (ICCRE), pp. 259-263, 2018.
[17] Y. Wang and Y. Cao, "Human peripheral blood leukocyte classification method based on
convolutional neural network and data augmentation," Medical physics, vol. 47, no. 1, pp. 142-
151, 2020.
[18] M. Sharma, A. Bhave and R. R. Janghel, "White blood cell classification using convolutional
neural network," Soft Computing and Signal Processing, pp. 135-143, 2019.
[19] J. Yao et al., "High-efficiency classification of white blood cells based on object detection",
Journal of Healthcare Engineering, 2021.
[20] J. Wu et al., "WBC Image Segmentation Based on Residual Networks and Attentional
Mechanisms," Computational Intelligence and Neuroscience, 2022.
[21] S. H. Rezatofighi and H. Soltanian-Zadeh, "Automatic recognition of five types of white blood
cells in peripheral blood," Computerized Medical Imaging and Graphics, vol. 35, no. 4, pp. 333-

475
 Sakarya University Journal of Computer and Information Sciences

Oguzhan Katar et al.

343, 2011.
[22] M. Mohamed, B. Far and A. Guaily, "An efficient technique for white blood cells nuclei automatic
segmentation," 2012 IEEE International Conference on Systems, Man, and Cybernetics (SMC),
pp. 220-225, 2012.
[23] O. Sarrafzadeh, H. Rabbani, A. Talebi and H. U. Banaem, "Selection of the best features for
leukocytes classification in blood smear microscopic images," Medical Imaging 2014: Digital
Pathology, vol. 9041, pp. 159-166, 2014.
[24] R. D. Labati, V. Piuri and F. Scotti, "All-IDB: The acute lymphoblastic leukemia image database
for image processing," 18th IEEE international conference on image processing, pp. 2045-2048,
2011.
[25] X. Zheng, Y. Wang, G. Wang and J. Liu, "Fast and robust segmentation of white blood cell images
by self-supervised learning," Micron, vol. 107, pp. 55-71, 2018.
[26] Z. M. Kouzehkanan et al., "A large dataset of white blood cells containing cell locations and types,
along with segmented nuclei and cytoplasm," Scientific reports, vol. 12, no. 1, pp. 1-14, 2022.
[27] K. He, X. Zhang, S. Ren and J. Sun, "Deep residual learning for image recognition", IEEE
conference on computer vision and pattern recognition, pp. 770-778, 2016.
[28] K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image
recognition", arXiv preprint arXiv:1409, 1556, 2014.
[29] A. G. Howard et al., "Mobilenets: Efficient convolutional neural networks for mobile vision
applications", arXiv preprint arXiv:1704.04861, 2017.
[30] J. Deng, W. Dong, R. Socher, L. J. Li, K. Li and L. Fei-Fei, "Imagenet: A large-scale hierarchical
image database," IEEE conference on computer vision and pattern recognition, pp. 248-255, 2009.
[31] [Online]. Available: https://keras.io/api/applications/ [Accessed in December 2022].
[32] R. R. Selvaraju et al., "Grad-cam: Visual explanations from deep networks via gradient-based
localization," IEEE international conference on computer vision, pp. 618-626, 2017.
[33] M. Yildirim and A. Çinar, "Classification of White Blood Cells by Deep Learning Methods for
Diagnosing Disease," Rev. d'Intelligence Artif., vol. 33, no. 5, pp. 335-340, 2019.
[34] A. Ekiz, K. Kaplan and H. M. Ertunç, "Classification of white blood cells using CNN and Con-
SVM," 29th Signal Processing and Communications Applications Conference (SIU), pp. 1-4,
2021.
[35] A. Girdhar, H. Kapur and V. Kumar, "Classification of White blood cell using Convolution Neural
Network," Biomedical Signal Processing and Control, vol. 71, 2022.
[36] S. Nahzat, F. Bozkurt and M. Yağanoğlu, "White Blood Cell Classification Using Convolutional
Neural Network. Journal Of Science," Technology And Engineering Research, vol. 3, no. 1, pp.
32-41, 2022.
[37] M. Ö. Karakuş and E. Özbay, “Lökosit Tespiti İçin Beyaz Kan Hücrelerinin Esa Kullanılarak
Sınıflandırılması,” Adıyaman Üniversitesi Mühendislik Bilimleri Dergisi, vol. 9, no. 17, pp. 333-
344, 2022.
[38] C. Jung, M. Abuhamad, J. Alikhanov, A. Mohaisen, K. Han and D. Nyang, "W-net: a CNN-based
architecture for white blood cells image classification," arXiv preprint arXiv:1910.01091, 2019.
[39] Z. Wang, J. Xiao, J. Li, H. Li and L. Wang, “WBC-AMNet: Automatic classification of WBC
images using deep feature fusion network based on focalized attention mechanism,” Plos one, vol.
17, no. 1, 2022.
[40] R. M. Roy and P. M. Ameer, “Segmentation of leukocyte by semantic segmentation model: A
deep learning approach,” Biomedical Signal Processing and Control, vol. 65, 102385, 2021.

476