Machine learning in diabetic retinopathy 

Dr Pearse Keane’s group at the UCL Institute of Ophthalmology and Moorfields Eye Hospital, as well as others, have shown the viability of health professionals to develop powerful AI machine learning models using automated platforms. These sophisticated modern models have incredible capability to identify patterns in data and perform certain tasks at superhuman levels.

Although these platforms don’t need extensive technical expertise or infrastructure, their inherent complexity makes them challenging for clinicians to understand and trust. In order for machine learning to be fully integrated into such a safety-critical domain as healthcare, there is a great need for systems that can explain their reasoning and conclusions to medical professionals.

This project takes one step towards addressing this need by validating a method that can automatically give an explanation of an AI machine learning model used for classifying diabetic retinopathy, as well as other retinal diseases. This model has been built and trained using existing datasets of images of the back of the eye, called fundus images.

The explanation will automatically connect known features of diabetic retinopathy to identifiers used by the AI model, thereby demonstrating its reasoning to the clinician.

The team hope to develop an automatic explanation of machine learning models to be used in conjunction with the automation of training machine learning models that they have already developed. Together, this will enable clinicians without extensive computational expertise or resources to not only train their own machine learning models, but also make sense of them. 

Ultimately, the team aim to empower any healthcare professional to safely and responsibly use machine learning to solve the problems in their specific field and improve patient care.