cheXrad

CheXrad is pulmonary application with 4 differents artificial intelligence models. It can be trained very fast with CoCoLo technology. The software’s goal is to support doctors’ decisions with analyzing x-ray images.

Each artificial intelligence network is trained separately so their knowledge can be specialized or generalized as well. The neural network analyze the content of an image independently from each other too. The result can be a simple decision about the X-ray picture: healthy - not healthy. In case of illness the concrete disease is also communicated.

The four separated neural networks work like healthcare counselor group of several specialists which only purpose is to support the doctor. It gives not only a second opinion but also third and fourth opinion of the actual case. The result of each network will be evaluated and visualized after the decision process. Multi-level hierarchy is the key so the computer could decide even if it did not see the actual problem during the training process.

The rate of false - true and of negative - positive results are important qualification indices about the quality of a model in healthcare researches. True negative means the examined case is negative in real life and the model also predicted negative result. In case of a true positive results, both the sample is positive (shows disease) and the network identified it positive. The result is false when the labeling and the reality is different. In case of false positive result the patient is healthy but the test resulted positive due to its accuracy limits. In case of false negative result the patient is sick but the test results negative. The most important goal of researchers is to reach the highest rate of true positive and true negative result while false result doesn’t show up at all or they appear rarely depending on the applied technology. Every test and method have their own accuracy rates.

The performance of our model on COVID-19 disease X-ray images are shown below:

One important note to the numbers: the available dataset is very slightly. We try to reach the best result during balancing between insufficient data and making a generalized model that performs quite well in a real-life situation.The exact training process depends on the goal, which type of result satisfies more the user’s expected desires and which type of result seems the most useful to the user. Focusing on real negative cases means the model assign people to the healthy group better. However, it may throw warning sign as a false positive label. Focusing on disease or illness means the computer are able to classify the real diseased patient very accurately with minor loss, but the border between healthy and not healthy is blurred and borderline cases may be selected with much higher loss.

Special condition of the fight against the COVID-19 disease is the slightlyness of available data. This circumstance highly encumber the construction of artificial intelligence based systems since neural networks need huge amount of data to train on. We solve and solved this issue by special methods. We clean the data with statistical and mathematical methods. Since only small amount of data is available we have to multiply it too. We found proper ways to do this. This method has its limits too. However it shows good accuracy we consider our COVID-19 model still to be in learning phase only.

The point of the whole system is to help doctors with information to make stable diagnose and to be able to make a fast and cost-efficient examination without virus-test. Therefore we do not want to replace doctors with machines. However, this is seen impossible in this time. On the contrary, we would like to disengage the heavily overloaded healthcare system by computing, analízing and ordering patient cases much more faster with computer than doctors do. It helps to the doctors and nurses to focus on those cases, upon which their contribution could not be ignored. This software gives the possibility of fast response and intervention to the doctors’ hand by making an ordered list based on the patients’ presumptive condition. Therefore the patients with bad condition could get medical aid and assistance faster, that increases the chance of recovery.