A bounty on the corona virus has been suspended on the Kaggle online community, which is owned by Google. Data scientists and AI enthusiasts are said to be attacking the virus.
A total of $ 10,000 is for the best Answering ten different research questions advertised. The basis for this is a data set compiled, among other things, by the White House, which consists of over 45,000 scientific articles on the subject of SARS-CoV-2 and COVID-19. On the basis of this data set, the data scientists are to answer, among other things, what is known so far about the transmission, the incubation period or the risk factors. There are already over 500 submissions. The first round of the competition runs until April 16.
Neural networks are said to diagnose COVID-19
But elsewhere, attempts are being made to use artificial intelligence in the fight against Corona. In a open database carry AI researchers from the University of Montreal, CT and X-rays of the lungs of COVID-19 patients together. Their goal is to use this data to train artificial neural networks that can diagnose the disease using pattern recognition. This has already been successfully tested in other areas of medicine. And also in the case of COVID-19 show up AI researcher from the University of California San Diego (UCSD) and the University of Birmingham behave confidently.
Both San Diego and Birmingham use artificial neural networks, more precisely: so-called convolutional neural networks (CNNs). It is a concept of machine learning based on the imitation of biological neural networks. Similar to a brain, CNNs traditionally consist of several layers of interconnected artificial neurons that process the data.
Birmingham announced with his Decompose, Transfer, and Compose (DeTraC) called model diagnostic inaccuracies of over 90 percent. However, according to the researchers from San Diego, who also achieve a very high level of accuracy, there is a catch: The amount of CT and X-rays available for training the CNN is still too small. This could lead to so-called overfitting, a feared effect in machine learning. The model works very well with the training data set, but loses a lot of accuracy with the much wider range of data in real use. The basic approach is promising. However, according to the researchers from San Diego, it is important for further research that a larger number of image data is available.
The researchers from Montreal want to contribute to this. They intend to set up an open database that is freely accessible to all scientists, but also because, according to reports appropriate software from commercial companies in China is already in use.
A reliable COVID-19 diagnosis using imaging methods would definitely be a major step forward. The previously common process based on the polymerase chain reaction requires a fully equipped laboratory; X-ray machines and other imaging methods are available in every hospital.