As the Covid-19 pandemic rages across the world, researchers and scientists as well as companies are increasingly using AI to assist in discovery of a cure and in understanding the disease plus the outbreak further.
Hyderabad based AI startup, Alpes.ai has been working on the detection of COVID-19 using Chest X-ray Images and here are their findings. If you have chest X-ray, upload it here and let us know how accurate the model is.
It was a few months ago when we first heard of COVID-19. We thought it would be controlled and won’t spread to countries. But after 2 months it has spread to almost all the countries. As rapid spread is the cause of this pandemic, early detection of the virus in humans can help in controlling the spread of virus from one human to another. The problem is the golden test for detection of COVID-19 the RT-PCR test is taking a lot of time to identify the person is whether COVID19 positive or negative. So, we thought of solving this issue from an AI perspective.
This dataset with standard pre-processing steps like resizing has been fed to a Transfer Learning-based VGG19 convolution neural network for feature extraction. Later these features are used to train the SNN(Sieve Neural Network) which is a research outcome of Dr. K. Eswaran and ALPES.ai, which is fast when compared to other Neural Network models. This algorithm separated each data point from other using planes. So, more specific information/pattern is recognised.
At first, we were in search of RT-PCT clinical data so that we can analyse the blood sample and detect COVID-19. But we weren’t successful in finding the data. Then we came across the chest X-ray images of COVID-19 affected patients. A Github repo is tracking all the images from across the world uploading to open source. Dataset – https://github.com/ieee8023/covid-chestxray-dataset.
This lead us to the idea of detection of COVID-19 from chest Xray Images.
But, to build the ML model we need data of non-COVID x-ray images too, then we found a dataset in Kaggle which has images of Normal patients and Pneumonia affected patients. (https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia). We combined these two datasets which resulted in a dataset of 3 classes (COVID-19, Normal, and Pneumonia).
This system can help in the detection of COVID-19 in a non-invasive way and can analyse the X-ray images in less than a second. This system is not a replacement to RT-PCR but, we are proposing a reversal in the present steps, where first RT-PCR is done and then X-ray. But, if we can reverse this process we can test many people through x-ray and then perform RT-PCR. As X-ray machines are portable, we can even test in remote places too. This system can be set up in Airports, Harbours, workplaces.
With a train set of 4,771 images, the SNN model is trained and tested on 1,193 images. This resulted in an accuracy of 98.07%, which is the new benchmark when compared to other models.
Our research work is available as a preprint on the research gate which explains in detail the approach of the model we built. Preprint – https://www.doi.org/10.13140/RG.2.2.19689.44645
An interesting approach by the team – help them show your upvote by sharing this within your community.
