The field of healthcare has been inching towards automation in recent years. In fact, this year itself has seen some groundbreaking studies – Predicting psychosis, detecting damage in knee joints, detecting cancer, among others. At the heart of these studies? Machine learning, of course.
Given the amount of data points that are generated each minute in this industry, it’s not surprising to see the amount of research being done. The latest comes from London-based DeepMind, and it tackles the ever-growing problems of eyesight.
Their AI system can interpret eye scans from routine tests with an accuracy that will blow your mind. The system is so advanced that it can recommend how patients should be referred for treatment for over 50 eyesight related diseases. And it does all this as accurately as renowned doctors. Score one for the machines!
In a detailed blog post, DeepMind’s researchers have laid down the approach they used to come up with their AI system. There are two neural networks are play, which are combined by the system. The first neural network (called the segmentation network) analyses the scan to come up with the features of the disease(s) it sees, like haemorrhages, lesions, irregular fluid, etc.
The second neural network (called the classification network), analyzes he results of the first neural network to present healthcare professionals with diagnosis and recommendations. Check out the below infographic DeepMind created to illustrate the difference between the current process and their own AI system:
The sheer number of scans that healthcare professionals process during the day is in the thousands and the aim of DeepMind’s AI system is to cut that down significantly. Not only does it automatically detect the features of eye diseases in a matter of seconds, it also prioritizes patients who need urgent treatment. Who said machines are bad for humans?
As always, DeepMind are at the forefront of breakthrough technology. But the next challenge for them is to convert this into a practical product which can be distributed to hospitals and clinics. As I mentioned above, their partnership with Moorsfield has yielded excellent results so that should encourage others to take it up.
Coming to the approach used, notice how they have made the entire system interpretable. There is no ‘black box’ associated with the neural networks used. The entire algorithm is transparent so that the results can be explained to patients. It’s truly a milestone research.
Useful