AI4COVID: Artificial Intelligence for the detection and prognosis of the disease CoVid-19
Today, the health systems are collapsing and it is not ruled out that there may be a second wave of CoVid-19 cases in Spain when containment measures terminates. Moreover, due to the scarcity of rapid diagnostic tests in our country, artificial intelligence can be a promising diagnostic tool that allows the rapid and early detection of sick patients. Therefore, the impact expected with the achievement of the project is a better control of potentially sick patients to focus efforts more efficiently where they are necessarily required. In this way, it would be possible to build a more solid base on the evidence of the impact of the disease in order to make health, socio-political and economic decisions.
The present project is intended to be approached from two different points of view (both based on deep learning). The first of these will be the creation of prediction models aimed at classification into different target classes (i.e. healthy X-rays, X-rays with COVID-19 X-rays and X-rays with other types of viral or bacterial pneumonia, etc.) by carrying out multiple groupings of these. Different classification and segmentation techniques of biomedical images will be carried out in order to identify and predict the different target classes. In addition, advanced concepts of deep learning will be employed such as knowledge transfer, parallel residual connections, design of new convolution architectures and implementation of loss functions according to the problem, among others techniques. As a second approach, the creation of a Content-Based Image Retrieval (CBIR) model based on auto-encoders is proposed. Given a new X-rays image, and by consulting the training database, it will show the most similar X-rays cases together with a series of population statistics to make decisions about the patient under study.
Based on the models developed, the aim is to create a diagnostic aid system that will make it possible to discriminate between the different lung conditions that patients may suffer from. The final objective is to develop a web platform that contains the multiclass prediction model integrated, so that it is possible to load a chest X-ray and predict instantly if it is a sample of a healthy patient, sick by CoVid-19 or by another type of viral or bacterial pneumonia, etc. The web application will also contain embedded algorithms capable of generating heat maps highlighting the areas of lung disease of interest so that it is very intuitive for expert personnel using the platform.