A deep learning model was developed which could identify if the person is suffering from Covid-19 or pneumonia or is a normal person based on the patients X-Rays.
COVID19 Classification Challenge https://www.kaggle.com/tawsifurrahman/covid19-radiography-database
• The numpy array was given had images of size 512x512 and local ram as well as RAM on Kaggle was unable to process due to lack of resources. • We resized the images to 256x256. • While trying to implement custom model with 10 layers the output accuracy was low. • There is a huge imbalance in dataset. We have very few images in covid19 class. • We performed augmentation for better results. • After performing augmentation, we got an accuracy of 92%. • But when we visualized the output through grad cam visualization. We figured that our model was detecting the alphabets R ,B1 etc. • Reading the literature of covid19 we came to a conclusion that the heatmap or features should be concentrated more on chest of the x-ray image. • So, we chose to use transfer learning. VGG19 was used for training. • Imagenet weights were used. • The output of the grad cam showed that the model detecting from the chest region of the image.
Results:
Model: Transfer Learning on VGG
VGG19 pretrained model was used in the experiment. Loss: categorical_crossentropy Optimizer: Adam Model weights with best validation accuracy was saved Epochs =40 Base Weights are taken from imagenet
Accuracy – 0.97198
precision recall f1-score support
Covid-19 0. 94286 0.97059 0.95652 34 Normal 0. 97642 0.96279 0. 96956 215 Pneumonia 0. 97235 0. 98140 0. 97685 215 Macro Avg 0. 96387 0. 97159 0. 96764 464 Weighted Avg 0. 97207 0. 97198 0. 97198 464
| precision | recall | f1-score | support | |
|---|---|---|---|---|
| Covid-19 | ||||
| Normal | ||||
| Pneumonia | ||||
| Macro Avg | ||||
| Weighted Avg |
Confusion Matrix
ROC value: {0: 0.98296853625171, 1: 0.9713551881946391, 2: 0.9786494816475203}
Output Visualization with GradCam (GradCAM-with-keras)
Heatmap of Covid
Heatmap of Normal
Heatmap of pneumonia
Training Loss
All necessary installs are written in the jupyter notebook. Tensorflow 2.0.2 Other requirements: RAM>= 16GB (if there is collision please contact us at sohampatil798@gmail.com)
• Augmentation of covid class images is needed to get good accuracy. • Images in the Dataset contains letters. • Smaller models learn these letters instead. • Through gradCam we can stop treating deep learning model like a black box. GradCam visualization helps in identifying if the features vector extracted at fully connected layer contains features which we need to analyze. • Grad Cam source link given in the readme doesn’t work with tensorflow >2.0. • We need to use this library.