The diagnosis of Covid-19 and other diseases has been accelerated thanks to an AI algorithm.
Thanks to a technique called Federated Learning, the University of Cambridge International Team has developed an AI algorithm that can be trained using the datasets of other hospitals and foreign facilities.
In this way, the diagnostics of Covid-19 and other pathologies are facilitated while keeping patients’ privacy intact.
FEDERATED LEARNING FOR DIAGNOSIS: MACHINE LEARNING WITHOUT CENTRALIZED TRAINING DATA
In contrast to the traditional learning techniques that describe automatic training on a single centralization of data (1), Federated Learning is a learning technique in which, in the algorithmic training phase, decentralized devices are used which however require machines or data without the need that these exchanged (2).
In this way the security and protection of the data is guaranteed.
See also Federated Learning to predict Oxygen needs.
THE ROLE OF DATA
In the early stages of the pandemic, many AI researchers devoted themselves to developing diagnostic models of Covid-19 even though many of these used low-quality data from “Frankestein” datasets.
The dataset on which the researchers of the International Team relied contained more than 9,000 CT scans of approximately 3,300 patients from 23 different hospitals, both in England and in China.
The framework outlined made it possible to develop reliable and accurate AI techniques in the field of medical diagnosis: “where previous models were based on arbitrary open source data, we worked with a large team of radiologists from the NHS and Wuhan Tongji Hospital Group to select the data, so we can start from a strong position, ”says co-author Hanchen Wang of the Engineering Department of the University of Cambridge.
Dr Michael Roberts of AstraZeneca and the Department of Mathematics and Theoretical Physics pointed out that every country has its own way of doing things, so you need as large a database as possible in order to create something that can be useful to most. of clinics.
In the article published by the researchers in Nature Machine Intelligence (3) it was also indicated how the scholars had to mitigate the biases caused by the different datasets and how they used the Federated Learning technique.
In this way they were able to train an Artificial Intelligence model capable of preserving the privacy of each data center.
Furthermore, the scholars have prepared a comparison between the AI models, validated by the researchers with the same data without superimposing them on the training ones, with the diagnoses elaborated by the team of radiologists on the same dataset, therefore the same CT used by the AI model, comparing the accuracy of the Artificial Intelligence model with the analysis of professionals in the sector.
CONCLUSIONS
The research conducted by the International Team of the University of Cambridge has developed an Artificial Intelligence model useful not only for the diagnosis of Covid-19 but also for many other diseases thanks to the analysis of CT scans from many different clinics, all guaranteeing the training of the model in a completely anonymous way while preserving the privacy of patients.
“By working with other countries, we can do much more than we can do alone,” Roberts pointed out.
Resources: