COVID-19: A Medical diagnosis using Deep Learning

Barney Darlington 19 Apr, 2023 • 6 min read

This article was published as a part of the Data Science Blogathon

Introduction

The ongoing global pandemic called COVID-19 is caused by SARS-COV-2, which was first found and identified in Wuhan, China, in December 2019. The lockdown in Wuhan and its neighboring districts failed to contain this virus which led to one of the worst humanitarian crises in the modern world, affecting millions of people all around the globe.

The virus spread quickly and mutated which gave rise to several waves of it, mostly affecting the third world and developing countries. The number of affected people is rising steadily as the world’s governments try to control the spread.

In this article, we will use the CoronaHack- Chest X-Ray Dataset. It contains chest X-Ray images and we have to find the ones that are affected by the coronavirus.

The SARS-COV-2 which we talked previously about is the type of virus that majorly affects the respiratory system, so Chest X-ray is one of the important imaging methods we can use to identify an affected lung. Here’s a side by side comparison:

So as you can see, how COVID-19 pneumonia can engulf the whole lungs and is more dangerous than both Bacterial as well as Viral types of pneumonia. I highly suggest you read the paper Transfer Learning for COVID-19 Pneumonia Detection
and Classification in Chest X-ray Images
which I’ve linked in the above image.

In this article, we will use Deep learning and Transfer learning to classify and identify X-Ray Images of lungs affected with Covid-19.

Importing Libraries and Loading Data

import matplotlib.pyplot as plt
import seaborn as sns
%matplotlib inline
import numpy as np
import pandas as pd
sns.set()
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import  *
from tensorflow.keras.optimizers import Adam, SGD, RMSprop
from tensorflow.keras.applications import DenseNet121, VGG19, ResNet50

import PIL.Image
import matplotlib.pyplot as mpimg
import os
from tensorflow.keras.preprocessing.image import ImageDataGenerator, img_to_array
from tensorflow.keras.preprocessing import image

from tqdm import tqdm
import warnings
warnings.filterwarnings("ignore")

from sklearn.utils import shuffle
train_df = pd.read_csv('../input/coronahack-chest-xraydataset/Chest_xray_Corona_Metadata.csv')
train_df.shape
> (5910, 6)
train_df.head(5)
Medical diagnosis using Deep Learning head
train_df.info()
info

Dealing with Missing Values

missing_vals = train_df.isnull().sum()
missing_vals.plot(kind = 'bar')
Medical diagnosis using Deep Learning missing value
train_df.dropna(how = 'all')
train_df.isnull().sum()
cat

train_df.fillna('unknown', inplace=True)
train_df.isnull().sum()
 Medical diagnosis using Deep Learning 1
train_data = train_df[train_df['Dataset_type'] == 'TRAIN']
test_data = train_df[train_df['Dataset_type'] == 'TEST']
assert train_data.shape[0] + test_data.shape[0] == train_df.shape[0]
print(f"Shape of train data : {train_data.shape}")
print(f"Shape of test data : {test_data.shape}")
test_data.sample(10)
 Medical diagnosis using Deep Learning shape
 Medical diagnosis using Deep Learning 2

We will fill the missing values with ‘unknown’.

print((train_df['Label_1_Virus_category']).value_counts())
print('--------------------------')
print((train_df['Label_2_Virus_category']).value_counts())
value.count Medical diagnosis using Deep Learning

Thus the Label 2 category contains the COVID-19 cases!

 

Displaying Images

test_img_dir = '/kaggle/input/coronahack-chest-xraydataset/Coronahack-Chest-XRay-Dataset/Coronahack-Chest-XRay-Dataset/test'
train_img_dir = '/kaggle/input/coronahack-chest-xraydataset/Coronahack-Chest-XRay-Dataset/Coronahack-Chest-XRay-Dataset/train'

sample_train_images = list(os.walk(train_img_dir))[0][2][:8]
sample_train_images = list(map(lambda x: os.path.join(train_img_dir, x), sample_train_images))

sample_test_images = list(os.walk(test_img_dir))[0][2][:8]
sample_test_images = list(map(lambda x: os.path.join(test_img_dir, x), sample_test_images))
plt.figure(figsize = (10,10))
for iterator, filename in enumerate(sample_train_images):
    image = PIL.Image.open(filename)
    plt.subplot(4,2,iterator+1)
    plt.imshow(image, cmap=plt.cm.bone)

plt.tight_layout()
 Medical diagnosis using Deep Learning tight layout

Visualizations

plt.figure(figsize=(15,10))
sns.countplot(train_data['Label_2_Virus_category']);
countplot

For COVID-19 cases

fig, ax = plt.subplots(4, 2, figsize=(15, 10))


covid_path = train_data[train_data['Label_2_Virus_category']=='COVID-19']['X_ray_image_name'].values

sample_covid_path = covid_path[:4]
sample_covid_path = list(map(lambda x: os.path.join(train_img_dir, x), sample_covid_path))

for row, file in enumerate(sample_covid_path):
    image = plt.imread(file)
    ax[row, 0].imshow(image, cmap=plt.cm.bone)
    ax[row, 1].hist(image.ravel(), 256, [0,256])
    ax[row, 0].axis('off')
    if row == 0:
        ax[row, 0].set_title('Images')
        ax[row, 1].set_title('Histograms')
fig.suptitle('Label 2 Virus Category = COVID-19', size=16)
plt.show()
virus category countplot

For Normal cases

fig, ax = plt.subplots(4, 2, figsize=(15, 10))


normal_path = train_data[train_data['Label']=='Normal']['X_ray_image_name'].values

sample_normal_path = normal_path[:4]
sample_normal_path = list(map(lambda x: os.path.join(train_img_dir, x), sample_normal_path))

for row, file in enumerate(sample_normal_path):
    image = plt.imread(file)
    ax[row, 0].imshow(image, cmap=plt.cm.bone)
    ax[row, 1].hist(image.ravel(), 256, [0,256])
    ax[row, 0].axis('off')
    if row == 0:
        ax[row, 0].set_title('Images')
        ax[row, 1].set_title('Histograms')
fig.suptitle('Label = NORMAL', size=16)
plt.show()
normal
final_train_data = train_data[(train_data['Label'] == 'Normal') | 
                              ((train_data['Label'] == 'Pnemonia') &
                               (train_data['Label_2_Virus_category'] == 'COVID-19'))]
final_train_data['class'] = final_train_data.Label.apply(lambda x: 'negative' if x=='Normal' else 'positive')
test_data['class'] = test_data.Label.apply(lambda x: 'negative' if x=='Normal' else 'positive')

final_train_data['target'] = final_train_data.Label.apply(lambda x: 0 if x=='Normal' else 1)
test_data['target'] = test_data.Label.apply(lambda x: 0 if x=='Normal' else 1)
final_train_data = final_train_data[['X_ray_image_name', 'class', 'target', 'Label_2_Virus_category']]
final_test_data = test_data[['X_ray_image_name', 'class', 'target']]
test_data['Label'].value_counts()
label counts

Data Augmentation

datagen =  ImageDataGenerator(
  shear_range=0.2,
  zoom_range=0.2,
)

def read_img(filename, size, path):
    img = image.load_img(os.path.join(path, filename), target_size=size)
    #convert image to array
    img = image.img_to_array(img) / 255
    return img
samp_img = read_img(final_train_data['X_ray_image_name'][0],
                                 (255,255),
                                 train_img_path)

plt.figure(figsize=(10,10))
plt.suptitle('Data Augmentation', fontsize=28)

i = 0


for batch in datagen.flow(tf.expand_dims(samp_img,0), batch_size=6):
    plt.subplot(3, 3, i+1)
    plt.grid(False)
    plt.imshow(batch.reshape(255, 255, 3));
    
    if i == 8:
        break
    i += 1
    
plt.show();
data augmentation
corona_df = final_train_data[final_train_data['Label_2_Virus_category'] == 'COVID-19']
with_corona_augmented = []


def augment(name):
    img = read_img(name, (255,255), train_img_path)
    i = 0
    for batch in tqdm(datagen.flow(tf.expand_dims(img, 0), batch_size=32)):
        with_corona_augmented.append(tf.squeeze(batch).numpy())
        if i == 20:
            break
        i =i+1


corona_df['X_ray_image_name'].apply(augment)

Note: The output was too long to include in the article. Here’s a small part of it.

augment
train_arrays = [] 
final_train_data['X_ray_image_name'].apply(lambda x: train_arrays.append(read_img(x, (255,255), train_img_dir)))
test_arrays = []
final_test_data['X_ray_image_name'].apply(lambda x: test_arrays.append(read_img(x, (255,255), test_img_dir)))
print(len(train_arrays))
print(len(test_arrays))
length

y_train = np.concatenate((np.int64(final_train_data['target'].values), np.ones(len(with_corona_augmented), dtype=np.int64)))

Converting all data to Tensors           

train_tensors = tf.convert_to_tensor(np.concatenate((np.array(train_arrays), np.array(with_corona_augmented))))
test_tensors  = tf.convert_to_tensor(np.array(test_arrays))
y_train_tensor = tf.convert_to_tensor(y_train)
y_test_tensor = tf.convert_to_tensor(final_test_data['target'].values)

train_dataset = tf.data.Dataset.from_tensor_slices((train_tensors, y_train_tensor))
test_dataset = tf.data.Dataset.from_tensor_slices((test_tensors, y_test_tensor))
for i,l in train_dataset.take(1):
    plt.imshow(i);
Converting all data to Tensors

Generating Batches

BATCH_SIZE = 16
BUFFER = 1000

train_batches = train_dataset.shuffle(BUFFER).batch(BATCH_SIZE)
test_batches = test_dataset.batch(BATCH_SIZE)

for i,l in train_batches.take(1):
    print('Train Shape per Batch: ',i.shape);
for i,l in test_batches.take(1):
    print('Test Shape per Batch: ',i.shape);
Generating Batches

Transfer Learning with ResNet50

INPUT_SHAPE = (255,255,3) 

base_model = tf.keras.applications.ResNet50(input_shape= INPUT_SHAPE,
                                               include_top=False,
                                               weights='imagenet')

# We set it to False because we don't want to mess with the pretrained weights of the model.
base_model.trainable = False

Now our Transfer Learning is successful!!

Transfer Learning with ResNet50

 

for i,l in train_batches.take(1):
    pass
base_model(i).shape
> TensorShape([16, 8, 8, 2048])

Adding a dense layer for Image Classification

model = Sequential()
model.add(base_model)
model.add(Layers.GlobalAveragePooling2D())
model.add(Layers.Dense(128))
model.add(Layers.Dropout(0.2))
model.add(Layers.Dense(1, activation = 'sigmoid'))
model.summary()
Adding a dense layer for Image Classification
callbacks = tf.keras.callbacks.EarlyStopping(monitor='val_loss', patience=2)


model.compile(optimizer='adam',
              loss = 'binary_crossentropy',
              metrics=['accuracy'])

Prediction

model.fit(train_batches, epochs=10, validation_data=test_batches, callbacks=[callbacks])
prediction
pred = model.predict_classes(np.array(test_arrays))
# classification report
from sklearn.metrics import classification_report, confusion_matrix
print(classification_report(test_data['target'], pred.flatten()))
classification

So as you can see the prediction is not half bad. We will plot a confusion matrix to visualize the performance of our model:

con_mat = confusion_matrix(test_data['target'], pred.flatten())
plt.figure(figsize = (10,10))
plt.title('CONFUSION MATRIX')
sns.heatmap(con_mat, cmap='cividis',
            yticklabels=['Negative', 'Positive'],
            xticklabels=['Negative', 'Positive'],
            annot=True);
confusion matrix

End Notes

This dataset was an interesting one, and the more I am learning Data Science and Machine Learning, the more I find this subject interesting. There’s soo many ways we can use data nowadays and using it can save countless lives. Thank you for reading this article. You can read more of my content at:

Barney6, author at Analytics Vidhya

The media shown in this article are not owned by Analytics Vidhya and are used at the Author’s discretion.

Frequently Asked Questions

Lorem ipsum dolor sit amet, consectetur adipiscing elit,

Responses From Readers

Clear

Yusef
Yusef 26 Jun, 2023

Great. How is transfer learning different from deep

Computer Vision
Become a full stack data scientist