As we gradually continue to delve deeper into the capabilities of artificial intelligence, we¡¯re finding more and more ways to apply it to medicine. Early warning systems after all is something AI is really good at, as evidenced by Google¡¯s latest creation.
Researchers from the University College London have been working with Moorfields Eye Hospital on Google¡¯s DeepMind AI, training it to identify dozens of common ocular diseases from 3D scans and recommend treatments. It¡¯s part of a multi-year collaboration between the three groups. The platform isn¡¯t ready just yet to be used in hospitals, but they indicate it could be in a matter of a few short years. DeepMind Health¡¯s head, Mustafa Suleyman, says the project could transform how we diagnose and treat patients with conditions that threaten their eyesight around the world.
The AI has been trained on 3D scans of patients¡¯ eyes, made using optical coherence tomography.? These OCT scans, which take about 10 minutes to make, involve bouncing near-infared light off the insides of a person¡¯s eye. This allows for a 3D image to be generated for doctors to study. In this case, the AI is being trained to understand the common patterns of data that appear in these OCT scans when the patient is suffering from an ocular disease.
The AI was trained on 15,000 scans of 7,500 eye patients treated at Moorfields sites, as well as the diagnoses they were given by human doctors. This way, the AI is taught to recognise the anatomy of the eye, as well as what it looks like when something is wrong, and the kind of treatment it deserves.
The team then tested the AI, comparing its diagnoses against those given by a panel of eight doctors. It passed with flying colours, recommended the same treatment as the human doctors over 94 percent of the time, for more than 50 different eye conditions.
It¡¯s a positive result to be sure, but medical professionals are obviously still wary. It¡¯s not that they¡¯re worried about being replaced, just that whenever a computer is making a decision in place of a doctor instead of just aiding their job, that there¡¯s room for fatal error. After all, even really extensive tests might convince us we know that the AI is working right, but a quirk of its training could just be coincidentally getting the right answers while going about the decision-making process all wrong.?
The simple scary fact is that we don¡¯t rightly know all the time why AI systems make the decisions they do. This is especially apparent in self-driving cars, and is much more pronounced in natural language systems too. Thankfully, DeepMind, UCL, and Moorfields researchers haven¡¯t forgotten this fact and are have included a features in their system to help minimise this possibility.
For one thing, the neural network¡¯s final decision is actually a product of multiple different algorithms that intersect. In addition, there¡¯s not just one diagnoses output, but instead several possible each associated with a confidence score given by the AI. That way, it allows overseeing doctors a chance to overrule the AI based on their experience.
More importantly, this software isn¡¯t intended for diagnosis, but more as a way to determine which patients are in more dire need of care. This way its diagnosis isn¡¯t taken as fact, but as an indicator of which patient should be examined first.
This way each of these features acts as a control valve for doctors to intervene. But the real test will only happen once it¡¯s actually put into service in clinics and hospitals.