Use of artificial intelligence(AI) in detection of COVID 19 case using CXR Data.

AI based Medical DiagnosticSys****tem

Detectio****n o****f COVID 19 through Chest radiography images

  In Collaboration with Nirma University and GSC Medical college

 Table of Contents

 Chapte****r 1 Abstract

Chapte****r 2 Problem Description

Chapte****r 3 Data

Chapte****r 4 AI Framework

 4.1 Classification network

4.2 Generating heatmap

Chapte****r 5 Model Details

5.1 Resnet50

 5.2 Proposed model

5.3 Heatmap generation

Chapte****r 6 Additional Experiments

Chapte****r 7 Deployment

Chapte****r 8 Future Scope

Chapter 1 - Abstract:

Rapid detection of severe acute respiratory syndrome coronavirus 2 (SARS CoV-2), which is responsible for coronavirus disease 2019 (COVID-19), using chest radiography images has life-saving importance for both patients and doctors. Collection of legitimate chest data along with labels is used to train a dense net model with a similar knowledge base whose results are further validated through severity heatmaps. We have had MOU agreement and ties with the Gujarat Cancer Society (GCS) which helps us by providing real world patient data so as to further improve our model classification accuracy.

Chapter 2 - Problem Description:

Due to the spread of the novel coronavirus, many lives are being lost as the virus is highly contagious. The most accurate test for Covid 19 in present is the RT-PCR test. But one disadvantage of this test is that it takes more than 24 hours for the results after the collection of the sample. This work felicitates a rapid and easy detection of the chances of a person having been infected with novel coronavirus by just analysis of the chest X- ray of the patient. By this at-least doctors can know if the patient should be quarantined until the results of his/her RT-PCR test or not.

Chapter 3 - Data:

For achieving the task of diagnosing COVID-19 from medical image analysis we collected and combined data from multiple sources and Additional Dataset was provided by the Gujarat Cancer Society

Final Dataset:

 

Norma****l

 COVID-19

| Trainin****g :

 7966

 973

| Testing:

 100

100

| To****tal:

 8066

 1073

To****tal Training Images: 8939 || Total Testing Images: 200

   Chapter 4 - AI framework:

  We propose the following pipeline for the detection of COVID-19 from chest radiography images. The pipeline consists of two main modules:

 4**.**1 Classification network:

We trained Resnet50, deep convolutional architecture with skip connections and identity blocks on the dataset  we created by combining several open-source datasets from scratch, for the purpose of classification.

 Evaluation:-

Test Accuracy: 85%

 To further improve the credibility of our results we propose transfer learning on the base model Che-x-Net which is trained on the base ChestX-ray14 dataset, which contains 112,120 frontal view X-ray images from 30,805 unique patients for detection of  14 diseases (mainly pneumonia). Since transfer learning allows us to use the knowledge gained from other tasks in order to tackle new but similar problems effectively, it can also aid our task. In our case, weights of feature extractor/convolutional blocks are kept the same as that of the Che-x-Net fully trained network. We remove the final output layer of the network and add a few dense layers followed by the output layer with 2 neurons (namely: Normal, COVID-19).

The loss function employed is categorical cross-entropy.

 Evaluation:

Test Accuracy: 92%

 4**.**2  Generating Heat map:

  For the ones classified as COVID19 positive our approach highlights evidence i.e. diseased patches in the radiography images for clinical users to ease their decision to accept or reject a deep learning-based chest radiography diagnosis.

 Chapter 5 - Model Details:

 5**.**1 ResNet50 (Residual networks):

A convolutional neural network that is 50 layers deep. It employs skip connections that mitigate the problem of vanishing gradients by allowing an alternate shortcut path for the gradient to flow. The ResNet-50 model consists of 5 stages each with a convolution and identity block. Each convolution block has 3 convolution layers and each identity block also has 3 convolution layers.

5**.**2 Proposed model - COVID19Net:

  COVID19Net is a 121-layer Dense Convolutional Network (DenseNet) pretrained on the ChestX-ray14 dataset and extended for the dataset created by us for COVID19 diagnosis. DenseNets improve the flow of information and gradients through the network, making the optimization of very deep networks tractable.The final layer is replaced by a few dense layers to be trained on our data. The final 2-neuron dense layer is followed by a softmax function for getting the classification probability.The network is trained end-to-end using Adam with standard parameters ( β1 = 0.9 and β2 = 0.999).We train the model using mini- batches of size 16. We use an initial learning rate of 0.001 that is decayed by a factor of  10 each time the validation loss flattens after an epoch, and pick the model with the lowest validation loss.

5**.**3 Heat map generation:

   We employ a prediction difference method for the visualization of trained models. We find the difference between the pixel values of the image patches in predicted images and image patches in the normal baseline radiography images. This approach generates a relevance score for each pixel which is visualized as a heat map.

Chapter 6 - Additional Experiments:

 1) Data Pre-processing:

As training data was collected from many different sources it was beneficial to pre- process all the images before feeding them into the network.

Here, we used 2 types of Data Pre-Processing methods:

A) Gamma - Correction

B) Image Histogram  Equalization

2. Use of Weighted Loss Function:

To overcome the high imbalance in the data-set which contained nearly 8000 normal images and only 1073 covid 19 images we employed a weighted loss function which gave more weight to the class with less number of data points so as to stop the network from being biased to one class.

3. Use of Balanced Data-Set:

To effectively train the model we also tried using random 1073 images from the Normal Class which resulted in the same number of images in both the classes ie Normal and Covid 19. By this, the model got trained on a perfectly balanced dataset.

A****fter all the Experiments the final Test Accuracy obtained is 97%

Chapter 7 - Deployment:

â—    To deploy the model in real-time a web application is prepared with the use of a flask framework. And the model is deployed on an Nvidia V-100 GPU.

â—   Additionally, a cropping function is provided for the user to extract the actual area of interest from the chest X-ray.

**Chapter 8 - Future Scope:**â—

With  increasing   size   of   accurate  data  the  model  should  be  constantly updated/trained to get an improving classification accuracy.

â—    Can also include CT-scans and severity segmentation maps for a holistic medical system development.

â—   To collect and store medical details of patients and suggest them possible check- ups and remedies on a timely basis.