The automated analysis of RX images using Artificial Intelligence (AI) and Deep Learning (DL) models has great potential for a rapid diagnosis of COVID19. A robust and accurate DL model can serve as a detection and support system for medical decision making. A set of recent studies claim to have achieved impressive sensitivities (> 90%), higher than those of expert radiologists (69%). These high sensitivities are due to bias in the most widely used dataset, COVID19 Image Data Collection. This set includes a small number of positive cases of COVID19, coming from very heterogeneous sources (at least 15 countries) and the majority of the cases are seriously ill, which drastically reduces their clinical value. To contrast non-COVID classes, they use RX images from various public repositories of lung diseases. The resulting models have no clinical value as they will not be able to detect patients of low and moderate severity, which are the goal of a clinical screening system. Faced with this situation, there is an enormous need for higher quality data sets constructed under the same clinical protocol and in close collaboration with expert radiologists.
Federated machine learning is an approach that enables DL models to be trained on-premises, in a distributed way, sent to a central server where their weights are added, and a consolidated model is forwarded to all devices / nodes for continue learning. It uses distributed computing concepts to keep track of each model in the nodes and add and update models in each of the nodes.
The overall objective of the project is to develop a robust system for the detection of COVID19 and bacterial and viral pneumonias for the triage of patients. It is divided into five specific objectives: