Covid-19 detection analyzing X-Ray images using a simple Convolutional Neural Net architecture

Just like all my articles, the opinions expressed in this article are of mine only as per my understanding and in no way represents anything that has to do with the organization I work for.

As several references mention,Chest X-Ray scans may be helpful to diagnose COVID-19 in individuals with a high clinical suspicion of infection. This small article is a demonstration of this using a simple deep learning architecture with a pretty reasonable accuracy, which is not perfect but makes a definite point in proving the effectiveness of AI as it may evolve in future. It is amazing to observe that with only a few lines of code, great results can be achieved!

The data that has been used for this experimentation, is publicly available in here:

https://github.com/shervinmin/DeepCovid/tree/master/data

Let's start by importing the required libraries as usual.

Created on Sat Nov 28 10:37:45 2020


@author: chakr
"""
import tensorflow as tf
from keras.preprocessing.image import ImageDataGenerator
import numpy as np
import glob

Next we need to preprocess the data.The ImageDataGenerator class is particularly useful and I would request the readers to try with different options to make the prediction even more accurate.

#data preprocessing
traingen = ImageDataGenerator(rescale = 1./255,shear_range=0.2,zoom_range=[0.2,1.8],horizontal_flip=True,vertical_flip=True)
training_set = traingen.flow_from_directory('covid19\\train',target_size = (256,256),batch_size=32,class_mode='binary')
testgen = ImageDataGenerator(rescale = 1./255)
test_set = traingen.flow_from_directory('covid19\\test',target_size = (256,256),batch_size=32,class_mode='binary')

Now, simply build the CNN model and train it with the training set.

cnn = tf.keras.models.Sequential()
cnn.add(tf.keras.layers.Conv2D(filters =32,kernel_size=3,activation='relu',input_shape = [256,256,3]))
cnn.add(tf.keras.layers.MaxPool2D(pool_size=(2,2),strides=2))
cnn.add(tf.keras.layers.Conv2D(filters =32,kernel_size=3,activation='relu'))
cnn.add(tf.keras.layers.MaxPool2D(pool_size=(2,2),strides=2))
cnn.add(tf.keras.layers.Flatten())
cnn.add(tf.keras.layers.Dense(units = 256,activation = 'relu'))
cnn.add(tf.keras.layers.Dense(units = 1,activation = 'sigmoid'))
cnn.compile(optimizer = 'adam',loss = 'binary_crossentropy',metrics=['accuracy'])
cnn.fit(x = training_set,validation_data=test_set,epochs=100)

Note the number of epochs in the snippet above. It usually takes 60 minutes to 90 minutes in my computer to finish the entire training for a 100 epoch run. So babysit the training in the beginning and based on the learning, you may adjust the same.

In my case, this is how the metrics look at the end of the run...

....
....
Epoch 99/100
21/21 [==============================] - 66s 3s/step - loss: 0.0645 - accuracy: 0.9759 - val_loss: 0.2242 - val_accuracy: 0.9381

Epoch 100/100
21/21 [==============================] - 67s 3s/step - loss: 0.0770 - accuracy: 0.9699 - val_loss: 0.0830 - val_accuracy: 0.9810

Now all that remains is the prediction for your data. You may feed single or multiple images in the model and collec predictions and check accuracy.

k = training_set.class_indices['covid']

def imageconverter(picture):
    import numpy as np
    from keras.preprocessing import image
    chk_img = image.load_img(picture,target_size = (256,256))
    chk_img = np.expand_dims(image.img_to_array(chk_img),axis = 0)
    return chk_img


imagelist = glob.glob('covid19\\val\\*.jpg')


chkimages = [imageconverter(i) for i in imagelist]


collector = []
for j in [imageconverter(i) for i in imagelist]: 
    predictions = cnn.predict(x = j)
    if predictions[0][0] == k:
        predictions = "Covid-19"
    else:
        predictions = "Normal"
    

    collector.append(predictions)
    

In my case, just without any optimization, in the first run, 88% accuracy was obtained in the validation images. I would encourage the readers to further optimize and share the results in the comments below.