A new artificial intelligence (AI) system can diagnose 50 different eye diseases as accurately as specialised ophthalmologists. The system is also able to determine which treatment is best for patients and which patients should be treated with priority.
The system has been developed by DeepMind, part of Google's parent company Alphabet, in collaboration with Moorfields Eye Hospital in the UK capital London. Tests with the system are yielding promising results. The system appears capable of diagnosing 50 different eye diseases as accurately as specialised eye doctors. Google stresses that these are early results, but states that they do show that the system is capable of diagnosing a wide variety of patients.
Time-consuming
Diagnosing eye diseases often involves the use of optical coherence tomography (OCT), an examination method that takes cross-sectional images of the eye. These scans need to be analysed, which is currently done manually by ophthalmologists. This makes performing OCT scans very time-consuming. At the same time, the need for such scans is great; around 1,000 OCT scans are performed per day at Moorfields Eye Hospital alone.
The large number of scans performed and the time-consuming analyses that need to be performed on them lead to a relatively long time between the scan and final treatment of a patient. This while, in some cases, patients need urgent care. DeepMind states that unexpected complications such as bleeding at the back of the eye can even cause patients to lose their vision.
Accelerating analysis
With the AI system developed in collaboration with Moorsfield Eye Hospital, DeepMind aims to provide a solution to this problem. The system should significantly speed up the analysis of OCT scans by automating this process. In doing so, the AI system can not only automatically detect abnormalities on the scans in just a few seconds, but also indicate which patients need urgent care. As a result, the AI system should help shorten the time between a scan and final treatment, so that fewer patients face vision loss.
The system uses two neural networks based on the human brain. Such a network is composed of several layers of chips that mimic neurons. Each layer makes an analysis of the data to be analysed and transmits its results to the next layer of neurons. This layer performs a new analysis on the transmitted data and passes the results of this analysis back to the next layer, after which the process is repeated. This process enables deep learning algorithm to learn independently and become smarter and smarter as they analyse more data.
Two neural networks
DeepMind's AI system uses two such neural networks. The first network is called a segmentation network by the company. This network analyses an OCT scan and maps the different types of eye tissue. The network also looks for clues that may indicate eye diseases, such as bleeding and fluid irregularities. The second neural network is used to classify the findings of the first network and provide both a final diagnosis and recommendation for treatment.
This recommendation is delivered as a percentage. This allows doctors to determine with what certainty the AI system has made its analysis and final diagnosis. This is critical, according to DeepMind, as ophthalmologists will always play a crucial role in diagnosing eye diseases. Thus, the AI system is only meant to support ophthalmologists and not replace doctors.
Also suitable for other scanners
The AI system has been specifically trained on the OCT scanners used by Moorfields Eye Hospital. However, DeepMind says the system can also easily be used to analyse scans with other eye scanners and make diagnoses based on them. Not only does this make the AI system usable worldwide regardless of the equipment used, it also allows the system to continue to be used in the future if OCT scanners are renewed or replaced.
However, the technology cannot yet be used in practice. DeepMind points out that the technology will have to be clinically tested and validated first. "However, we are confident that this system could eventually transform the diagnosis, treatment and management of eye diseases," DeepMind said.
Author: Wouter Hoeffnagel
Source: DeepMind
Source: Nature Medicine
Source photo: Pixabay / TobiasD
