CNN-SVM Learning

Approach Based For White

Blood Cell Classification
By Vijay Shankar Guide:Dr.V.V Sowmya
(CB.EN.P2CEN19015)
Types of White Blood Cells
Eosinophil

Monocyte

Eosinophil
Lymphocyte

Neutrophil
A more Clear Representation
●
 Eosinophils

● They are the cells of leukocytes, which are present in the immune system.

● These cells are responsible for combating infections in parasites of

vertebrates and for controlling mechanisms associated with the allergy and
asthma.

● Eosinophil cells are small granulocyte, which is produced in the bone marrow
and makes 2 to 3 percent of whole WBCs. These cells are present in high
concentrations in the digestive tract.
 Lymphocytes

● They play a vital role in producing antibodies.

● They are commonly known as natural killer cells.

● They play an important role in body defence.

● These white blood cells are colourless cells formed in lymphoid tissue, hence referred to as
lymphocytes.

● There are two main types of lymphocytes – B lymphocytes and T lymphocytes.

These cells are very important in the immune systems and are responsible for humoral and
cell-mediated immunity.
 Neutrophils

● They are normally found in the bloodstream.

● They are predominant cells, which are present in pus.

● Around 60 to 65 percent of WBCs are neutrophils with a diameter of 10 to 12 micrometres.

● The nucleus is 2 to 5 lobed and cytoplasm has very fine granules.

● Neutrophil helps in the destruction of bacteria with lysosomes, and it acts as a strong oxidant.

● Neutrophils are stained only using neutral dyes. Hence, they are called so.

● Neutrophils are also the first cells of the immune system to respond to an invader such as a
bacteria or a virus.
 Monocytes

● These cells usually have a large bilobed nucleus, with a diameter of 12 to 20 micrometres.

● The nucleus is generally of half-moon shaped or kidney-shaped and it occupies 6 to 8 per

cent of WBCs.

● They are the garbage trucks of the immune system.

● The most important functions of monocytes are to migrate into tissues and clean up dead
cells, protect against the bloodborne pathogens and they move very quickly to the sites of
infections in the tissues.

● These white blood cells have a single bean-shaped nucleus, hence referred to as Monocytes.
Problem Statement
● The microscopic inspection of blood smears provides diagnostic information
concerning patients’ health status.
● The inspection results of the differential blood count reveal a wide range of
important hematic pathologies.
● The presence of infections, leukemia, and some particular kinds of cancers
can be diagnosed based on the results of the classification and the count of
white blood cells.
● For Example when you have too much Eosinophil you can get Eosinophilia
which is a chronic illness which damages tissues.
● Traditional Counting process is tedious and pretty slow.
Existing Work
● This is a challenging data set in the sense that the White blood cells need to
be separated from the Red Blood cells.
●

W-Net: A CNN-based Architecture for White Blood Cells Image Classification(Changhun Jung et al 2019)
WNET
Dataset description ( Shenanigan 2019)

● Dataset taken from Kaggle , original credits to Shenanigan

● This dataset contains 12,500 augmented images of blood cells (JPEG).
● There are approximately 3,000 images for each of 4 different cell types
grouped into 4 different folders (according to cell type).
● Original images are only 410 images (pre-augmentation)
● There are approximately 3,000 augmented images for each class of the 4
classes as compared to 88, 33, 21, and 207 images of each of the classes.

https://www.kaggle.com/paultimothymooney/blood-cells
Segmentation Approach

● Colour based segmentation using

● They have achieved an accuracy of 99.1 percent using
segmentation
● This was tried against 450 images.
● CellaVision Competency Software Databases
● Multi-layer Perceptron achieved the highest
performance

A Neural-Network-Based Approach to White Blood Cell Classification(Mu-chun-su et al.)

Segmentation
● Segmentation is facilitated by the different colour present
● Normalization of the image results in separation of color components which
facilitates segmentation here
Histogram

Smaller section
indicates the Blue
region with patch
Which is the Region
under consideration.
So Normalization
clearly helps in
separation
CNN for Transfer Learning

CNN-SVM Learning Approach Based Human Activity

Recognition(Hend Basly et. al.)
 Source: Kaggle segmentation for moon rock
Transfer Learning
PROCESS

Random Forest

Images
INPUT

Transfer Learning
Support Vector Machines
(VGG 16)

Augmentation CNN ResNet Inception V2

Feature extraction
● The input images are resized from 240x320 into 100x100 (
RGB)
● In order to make it compatible with the output architecture the
dimension is expanded by 1 [1,100,100,3]: Tensor
● The VGG16() pretrained model is downloaded and the last
layer is removed
● This leaves the 4096 fully connected layer as the output
● This is resized to 64x63x3 and extrapolated to 100x100x3
Random Forest
● Gini is used to measure the quality of the split in contrast
with entropy
● Maximum depth of the tree is chosen arbitrarily until they
can not be split further
● The Tree is initiated with class weight one ( default)
● The number of trees in the forest is chosen as 100.
Support Vector Machine
● The Hyper planes are separated using the RBF
kernel
● Shrinking Heuristics is used to remove the bounded
elements
● No limit set on number of iterations
● The method of classification is the one vs Rest
methodology
Inception Resnet V2
● The method followed here is to freeze the final first 70 layers and train
the remaining 30 layers of the architecture.
● Image Augmentation is used sending 30 images per batch.
● A total of 2000 epochs are thus run.
● Considerable ease has been observed using this method , the training
time is much lower and the training accuracy is also very high
● This work aims at understanding the efficiency of Neural Network as a
Feature extraction , when the image features are very small
Deep Learning Results
Deep Learning Report
Support Vector Machine
Random Forest
SVC POLYNOMIAL