Blood Cell Image Classification: CNN and Transfer Learning with Python

This report delves into the application of Convolutional Neural Networks (CNN) and transfer learning, specifically the Xception model, for the classification and extraction of blood cell images. The study investigates how CNNs can be employed to analyze microscopic blood images, distinguishing between various cell types such as red blood cells, white blood cells, and platelets. The methodology involves using Python for image processing and creating a deep learning architecture to automatically extract features. The report highlights the efficiency of CNNs in identifying structural, color, and textural characteristics within blood cell images, supporting medical diagnosis. Furthermore, it explores the use of pre-trained models to enhance the accuracy of classification, demonstrating the benefits of transfer learning. The research also compares the CNN approach with other classifiers to determine the best success rates, which is mainly obtained by Quadratic discriminant analysis. The report also examines the use of automatic thresholding and adaptive contouring for cell segmentation, addressing potential errors in blood cell extraction. The goal is to develop an automated, high-throughput method to improve medical diagnostic capabilities and predict actual outcomes.

Extraction of blood cell image classification using convolution neural network
and transfer learning pre-trained model by using python
1

ABSTRACT
It has summarized about the concept of blood cell extraction which become easier to divide into
different multilayer cells. It has been identified that acquisition of image that could be arranged
in short terms and also wide spread stress fibres as long bundles. Convolutional neural network
technique, which composed by pooling layers that support in blood cell classification. The
primary research project is to generate the specific features from image processing. In order to
identify the extraction of blood cell image the report provide the learning about the hierarchical
feature, means that computed by multi layers. It consists of large number of data sets but it easily
classified with the help of convolutional neural network. The concept related convolutional
neural network because it help for identifying structure, colours, texture of blood cell image. In
term of medical, it is the most efficient method extraction of blood cell image classifications.
2

Contents
INTRODUCTION...........................................................................................................................3
LITERATURE REVIEW................................................................................................................9
RESEARCH METHODLOGY.....................................................................................................15
DATA INTERPRETATION.........................................................................................................21
RESULT AND FINDING.............................................................................................................37
CONCLUSION..............................................................................................................................39
RECOMMEDNDATION..............................................................................................................43
REFLECTCTION..........................................................................................................................45
REFERENCES..............................................................................................................................50
3

Title: Extraction of blood cell image classification through convolution neural network and
transfer learning (pre-trained) model by using python.
INTRODUCTION
The extraction of blood cell image classification can be possible with the help of
convolutional neural network. it can be defined as manipulating an image to achieve standards
that categorized into different images. It is well known method in term of blood cells
classification which may include red blood, white blood and platelets (Sahlol, Kollmannsberger
and Ewees, 2020). During blood extraction, leucocyte plays important role in term of human
immune functionality. Generally, it mainly identified the hematologists that can use for purpose
of granulated information and size or shape in leukocytes. It can be divided into different blood
cells such as basophil, eosinophil and non-granular cells. It often uses the basic information on
the basis of categorizes therefore, research study of blood cell classification which has important
in term of medical diagnosis.
The microscopic blood image of every patient play important role for purpose of diagnosis
and also managed with many diseases. In blood cells, there are various type of cells such as
white blood cells and red blood cells. This type of blood cell play important role in human body
which maintain immune system and fight against with viruses, bacteria and microbes. These are
extremely 5000 to 10000 WBC in adult’s human body, having values outside in particular
specified manner. It has increased to multiple diseases such as anemia, deficiency disease. That’s
why, primary reason for using convolutional neural network for blood cell image classification.
In this way, it can easily identify any problem that occurs in the human body (Coulibaly, Kamsu-
Foguem, Kamissoko and Traore, 2019). It is consider as modern technology that always support
for collecting large information about blood cells. Nowadays, the study was conducted for the
extraction of blood cell image classification. In this study, it mainly focused on the convolutional
neural network and use as a model where it will create deep learning architecture. It also used in
the different classification process of generate blood cell images. Main feature is that when it
automatically extracting large amount of cells and successful provided through architecture. As a
result, it is obtaining through multiple classifier on the blood cell architecture and also compare
with other one. Sometimes, it will identify the best success rate which mainly obtained by
Quadratic discriminant analysis. On the other hand, various type of classifier along with
4

extracted attributes and also increase success rate. WBC microscopic blood image accessible
from the convolutional neural network. In order to adopt and compare with other one.
A convolutional neural network is based on technique that help for identifying data in the
form of images. It can take action as input while assign into different aspects that able to
differentiate one from another (Coulibaly, Kamsu-Foguem, Kamissoko and Traore, 2019). The
process must be used the pre-processing which require for filtering the characteristic so that it
become easier to identify in proper manner. The entire architecture of CNN is analogous that
establish the connectivity in pattern of neuron in human brain. Recently, Convolutional neural
network has been widely implemented in terms of image classification (Jokhio and Jokhio,
2019). In order to achieve the unprecedented result or outcome. In this way, it is significantly
improving the overall performance of different medical imaging medium or applications. CNN is
highly efficient in terms of image processing, since features may be used in different images and
generating the output. It also demonstrates the optimization process which consider specific
parameters. Therefore, it has performed to minimize the difference between ground label in
blood images and generating results.
It is the best way to show the clearance in the performance of cell image while it can be
classified into proper manner. Apart from that Transfer learning mainly refers to the migration of
training model which consists of multiple parameter in the model. In this way, it helps for train
new model while identifying different layers in the image classification (Umapathy and et.al.,
2019). It is primarily used as base line of network that contain specific layer with randomized
determine the weights. It became useful for facilitating the transmission features in the same
layer as Xception which transferred in the similar model. In additionally, Xception is basically
used to categories the RGB channel which may consider as input and take single channel map.
Multiple layers in cell image classification, they are connected with each other and create as
single data set. Afterwards, it can be used the original learning rate to easily train entire layer of
blood cell.
In research study, it has proposed the Convolutional Neural Network method that will
help for classifying the blood cells and shown in the form of images. It has recognized the
classification of cells with rate of 94.2%. it will be used 100 images to identify their unique
properties. Thus, it implements the automatic threshold and adaptive contour for segment cells. It
helpful for identifying the errors if in case it doesn’t found the extraction in blood cells. The
5

recognize rate of blood cell is becoming higher because of Convolutional neural network which
classified into different manner (Coulibaly, Kamsu-Foguem, Kamissoko and Traore, 2019).
Therefore, it is implementing an automated and high throughput blood cell classification method,
it could be enabling technology to improve overall medical diagnosis capabilities, predicts actual
outcome and result. The primary objective of research project is to use CNN method for
classifying the blood cell image.
The report will discuss about the effectiveness of convolution neural network and transfer
learning model that easily monitored the activities within blood cells. The technique will help for
recording large number of data which called pre-processor data set. Sometimes, it automatically
turns into the 12,444 blood cells and enhanced towards image classification. With the
improvement in the overall performance of computer through convolutional neural network,
which has become most popular technique that record image of classification. Here, it become
easier for researcher to understand structure (Qiu and et.al., 2019).
In this report, it will propose the architecture that mainly combined with CNN
(convolutional neural network). It also integrated with the features from CNN and perform the
activities of blood cell classification (Coulibaly, Kamsu-Foguem, Kamissoko and Traore, 2019).
On the other hand, CNN must be used the pre-trained model during dataset of image collection
and retain its overall parameters. The pre-trained CNN model is mainly used for purpose of pre-
processed training which considered as input that extract and save obtained features
(Loganathan, 2019). The CNN model can perform the activities and apply multiple size and
weight matrix. In order to generate various features maps in processing. The proposed blood cell
classified by technique or method which contribution towards result. The research study will
identify the advantage of convolutional neural network and transfer learning in the extraction of
blood cell image classification by using python programming.
Background study
Researches about the extraction of blood cell classification in recent years, based on
Convolutional neural network technique. it is the most common approach that applicable image
processing algorithm for consists of different steps such as morphological, clustering,
segmentation, extraction, evaluation and classification. It also adopted the machine learning
technique in order to detect and classify blood cell images. Furthermore, Convolutional statistical
feature such as correlation that are extracted for give input in machine learning. In mostly, it
6

should be needed to decide about the features that must be utilised in order to increase blood cell
image classification accuracy. For instances, Transfer learning will help for extract high level
features which are automatically perform at same time. Therefore, it has proposed the
convolutional neural network architecture to determine abnormal and normal blood cell images.
In past decade, CNN is considered as technique which help for reducing process time and allow
to skip different steps. In this way, it can easily identify better result or outcome.
The segmentation process can help for detecting the blood cell classification image while
identifying background, plasm of blood image by using image processing (Coulibaly, Kamsu-
Foguem, Kamissoko and Traore, 2019). It is based on size, shape, colour and segmentation. It
has been proposed and combined with other technique for detecting and segmenting. Extraction
features representing as important step to classify the area, compactness, multiple layer and
distribution. Through existing theories, it works have done and also classified in context of
feature extraction. Furthermore, evaluation process can be implemented during classification
which help for generating numeric metric such as accuracy. It is the most important concept for
identifying overall efficiency and performance. The prediction may be classified into positive
and negative ways. During convolution neural network technique, it should be considered
different parameters that mainly define as testing and training protocols.
In the process of implementing the machine learning technique in term of blood cell
image classification, identifying meaning features that representation lies at core of success rate
into generate desirable outcome. The Image analysis by using computer based analysis and also
used the software technique with hand-engineered features (Sahlol, Kollmannsberger and Ewees,
2020). Sometimes, it also supported in the decision making where in case if it has found any
problem within blood cells. In past decade, demand of process has been identified where
expertise in analysing overall angle, size, position and back ground of image. In order to resolve
the challenges by devising hand-engineered features that easily capture variation into specified
manner.
Deep learning model is primarily used to cascade of multi layers, representing the raw
information or data. It applied high level features that supported to abstract from lower layers
and also helping for decision making. As a resulting, it acquire end to end features and classified
blood cell image in different manner. During image recognition, it is an essential source of
various data which mainly lies in the relationship among pixels. CNN, it would be created the
7

classes of model and also designed to exploit distribution information by using receptive fields,
pooling layer and shared weights. The ImageNet architecture is representing the visual image
and easily boosted the overall performance of convolutional neural network. In another way, it
will be classified natural image of blood cells in proper manner. There are several CNN such as
ResNet, VDDNet and GoogLenet that are demonstrated significantly improvement in the
challenge and issue. Another important model such as Xception, which was proposed that help
for depth analysis. In order to identify separable convolutions. This type of model is applicable
in the extraction of blood cell classification because it can outperform inception on the imageNet
during data classification (Sahlol, Kollmannsberger and Ewees, 2020). Convolutional neural
network has been categorised into variant which are densely connected with network system. It
also utilised the entire network architecture which are supposed to establish connection between
different layers. The imageNet model has achieved the improvement plan over start of art while
enhancing the computation action through different parameters. It also promising the overall
performance of convolutional neural network which accompanied after checking availability of
large amount of data or information. Sometimes, it has increased uncertainty due to medical
imagery transfer learning.
On the other hand, it applicable pre-trained model which ether find tuned on the specific
data, extracted to recognition task (Sahlol, Kollmannsberger and Ewees, 2020). The model has
gained the knowledge about the entire functionality of CNN and then afterwards, it can examine
the generic features from large datasets such as underlying operation through ImageNet model.
In previously, it has learned the skill to improve situation or condition in generalised manner. In
this way, it has recognised the CNN trained on the dataset that could be served as extraction
feature during blood cell classification through computer vision tasks. It also improved the
overall performance and also compare with different methods.
Transfer learning
It is the most popular method or technique in term of computer vision because it allows
us to build the accurate models. By using transfer learning, it is mainly starting with learning
process that automatically scratch and start from different patterns. It have been learned when
resolving the problem and issue. It became leverage previous learning about convolutional neural
network for calculating the different multilayers of blood cell images. Sometimes, it also
undertaken as deep learning version of expression (Srivastava and et.al., 2019). During computer
8

vision, transfer learning is expressed through use of pre-trained models. This type of model was
trained on the large benchmark dataset to resolve problem in proper manner. Accordingly, it can
be used the computational cost of training such as models which become common practices to
import or export. A comprehensive review the overall extraction of blood cell image and identify
the entire performance of trained models.
Convolutional neural network
It is based on the pre-trained model which mainly used as transfer learning on large scale.
In particularly, CNN has been shown into excel format in wide range of computer vision. It has
performed the different task to identify accurate result or outcome (Islam, Mubassira and Das,
2019). It is higher performance and easiness during training. In another way, it is to be
considered the main factor which driving the popularity of Convolutional neural network. A
typically CNN has divided into two different parts.
 Convolutional base, which has composed by stack of convolutional and other type of
pooling layers. In this way, it can easily identify primary goal is to generate specific
feature from blood cell image classification.
 Classifier, it has composed with multilayers which are established a strong connection
between them. The primary goals of classifier is to divide into small image blood cell and
easily detect the features. On the other hand, A fully connected with different layers
where neuron have establish connection with proper activation.
One important aspects of deep learning model is that when they can automatically learn entire
hierarchical features. So as handle to be computed features by multilayers. Generally, it can be
reused in the different domain while computed by last layers. The entire process may depend on
the data sets and various task. As a result, convolutional base of network especially from lower
to higher layers (Kim, Moon and Kang, 2019). During task execution, it is generally referred the
specific features whereas classifier part. Furthermore, it has pre-trained model for fulfilling
specific need and requirement. It will be starting to remove its original classifier and then add
new classifier that fit for specific purpose.
For image classification, it is standard approach to use the multiple stack of fully
connected layers which mainly followed by Softmax activated layers. It is helping in the output
layer where distribution of probability related result generation. It can be shown the possible
class label and then just require to classify image as per probable classes. Another type of
9

approach such as Global average pooling, it is based on the approach which are established the
connection with different layers on convolutional base. Sometimes, it is adding the pooling layer
and feed its outcome but it directly associated with the softmax activation layer. Linear support
vector machine is a type of approach which can improve the classification accuracy with the help
of training linear. It is mainly classified on the feature that will be extracted through
convolutional neural network.
Aim:
To investigate the Convolutional Neural Network and transfer learning to classify the
images by using python programming language.
Objective:
 To determine the concept of convolutional neural network and transfer learning.
 To identify advantage of convolutional neural network and transfer learning in the
extraction of blood cell image classification.
 To investigate the classification of blood cell image by using python programming
language.
Research Questions:
 What is concept of convolutional neural network and transfer learning?
 What are different advantages of convolutional neural network and transfer learning in
the extraction of blood cell image classification?
 What are classifications of blood cell image by using python programming language?
LITERATURE REVIEW
Theme:1 Determine the concept of convolutional Neural network and transfer learning.
Qi and et.al. (2019) A convolutional neural network is based on the learning technique
which mainly used to take input image and assign significant to different aspects in blood cell
images. In this way, it able to differentiate one from another. Sometimes, it is also known as pre-
processing which required for handling quality of image. While another method is to filters with
training convolutional has an ability to identify characteristics. CNN method can be categorised
into deep learning while identifying the difficulties in neural network which may overcome
through pre-training. Afterwards, it can be implementing layer wise pre-training in order achieve
the specific results or outcome. Convolutional neural network is a kind of network which help
for designing an effective structure effectively reduce the complexity in image recognition. This
10

type of technique will may include specific features such as adaptability, simple structure design
and pre-training parameters.
Generally, the structure of CNN may include two different layers, which is one of most
common features of extraction. Therefore, each neuron considers as input that connect with local
fields and extract from local features. The feature map is mainly used as sigmoid function as
activated functionality within convolution network and make features to shift invariances. CNN
is primarily used to determine the displacement and in another form of design two-dimensional
graphics. It is the most effective features that can help for identifying detection layer of
convolutional neural network by training information.
In this way, it plays important role in blood cell image classification where convolution
neural network technique uses to extract feature from raw blood cell image. The overall
architecture of CNN may include different layers such as fully connected layer, convolutional
layer and polling layer. In most of cases, convolutional layer computer with the neuron outcome
by estimating overall weight of input, adding bias to weight sum and then applying to filter
effectively. In this way, it is considered as one of important activation function with CNN
technique whether neuron should be activated. On the other hand, pooling layer will change to
reduce overall size image of blood cells. In this way, it will reduce the parameters number as
well as computations. The fully connected layers are containing a lot of neuron which directly
related to activation function from convolutional layers. In this process, it must use network
containing different layers which performed extraction while classified overall features. The size
of input is perfect to determine each layer in proper manner. The conventional layer is applying
filter option to categorise in different manner. In this way, it can be identified that it become
simplest technique to enhance level of contrast image and detecting overall distribution of blood
cell image pixels. In order to stretch the range of intensity value and calculate desirable outcome.
Byra and et.al (2019) argue that convolutional neural network is one of most common
approach in term of deep learning structure for image classification. CNN network become
popular after exceptional performance and identifying actual blood cell image quality. In
additional, it is similar to the other forward network in the input which are feeding through input
layer and processed. Moreover, Convolutional neural network has different layers which may be
including input which allows for capturing representative features from input data. Another step
11

of CNN concept is to utilise identical entire weights to process specific input and easily
synthesises its unique features.
Fahimi and et.al. (2019) transfer learning is based on the machine learning technique
where it can be developed model for reusing task in term of blood cell image classification. It
can be starts from initial step to last step. It is to be considered the most appropriate approach in
term of deep learning where it will be used pre-trainer model. It is mainly starting from hardware
vision where how they can recognise image through natural language processing. It is vast topic
which compute with resources to develop network model for identifying any problem through
image recognition. At that time, it can apply transfer learning technique to transfer entire weight
parameters that covered pre-trained on image recognition. It should be collecting large amount
data set which adopt to customise loss function. It allows for neural network to train and cover
faster speed in image recognition, classification. Hence, this process is mainly applied to detect
abnormal condition in cell which by means of segmentation as well as classification. As a result,
it has shown the 89% accuracy that’s why, method always proposed for further enhancement by
taking as a classifier. Yin and et.al (2020) said that transfer-based learning cell image
classification obtain the multiple patches which are exhausted in pixel wide sliding. However, it
also identified the major drawback whereas splitting the application on the basis of blood cell
classification. It will be generating large number of unwanted patches in different layers.
Although transfer learning and convolutional neural network have shows as good result
or outcome in terms of blood cell classification. Sometimes, it can be divided inti white blood
and red blood cells. It is very important to understand functionality of human immune cells and
then after it can be used method to divide range of blood cells into different manner.
Theme: 2 identify advantage of convolutional neural network and transfer learning in the
extraction of blood cell image classification.
Coulibaly, Kamsu-Foguem and Traore (2019) Convolutional neural network is now go
beyond every cell image and also relates to the specific problem in blood cells. In this way, it is
successfully recommending the CNN technique to process in blood cell image classification as
well as recognition. The advantage of convolutional neural network is that when it automatically
detects the features without supervision of human. It is the most powerful computational
technique which are becoming efficient to identify any type of problem within blood cells. It
benefits of special convolutional and also pooling multiple operations, it is performed different
12