2. The new coronavirus (COVID-19) is an
acute, deadly disease that originated in
December 2019 and spread globally from
Wuhan Province, China. The epidemic of
COVID-19 has been of great concern
to the medical community because no
efficient cure has yet been found.
Real-time reverse transcription polymerase
chain reaction(R T -PCR) test has been
described by the World Health Organization
(WHO) .
Introduction
3. Problem
The spread of the Corona virus in the world,
its transmission from a human to a human
being , and the inability of everyone to
perform a diagnostic scan and resort to
chest rays to detect the presence of
pneumonia or the presence of the Corona
virus.
12. Pre-Processing
After capturing the X-Ray images, we are applying
the preprocessing techniques on digital images
like RGB to Gray scale conversion and used
appropriate filtering techniques.
13. Image Enhancement
The first instance of the input X-ray scan is enhanced
using the techniques explained below :
1) Median Filter
2)Fuzzy Histogram Hyperonization.
15. Stacking and Augmentation
- Image stacking: The purpose of Image Stacking is to
move the individual segment images so that they fall
precisely on top of each other.
- inbalancing..The original dataset (without augmentation)
contains only 554
chest X-ray scans of COVID-19
17. Training &
Classification
In this study, deep learning was used for classifying images into
COVID-19 or Normal categories. An ensembled model was built by
concatenating the features of three different Convolutional Neural
Networks
18. VGG-16
VGG-16 is a simple 16 layered
Convolutional Neural Network . It has
convolutional filter of size 3 × 3 and
pooling filter of size 2 × 2.
19. ResNet50
ResNet-50 is a residual network with 50 layers
stacked together and has shortcut or skip
connections. The skip connection passes the
same information in the network. Passing the
same information allows that the model does not
degrade by losing the information
20. MobileNetV2
MobileNetV2 is a very light, low-latency and low-
powered model which requires very low
hardware setup for training a model. It has linear
layers for linear bottleneck between the layers
which prevents non-linearities from destroying
the information
21. Ensemble Model
The flatten output features from these sub-models are then concatenated
to make an ensemble of these three models. A meta learner is created
for classifying these features in one of the three categories .
22. Evaluating the performance
For evaluating the performance on test set, four evaluation metrics accuracy, precision, recall,
F1- score and are derived from the confusion matrix. The formulas for these metrics are given below