Paolo Silva, MD: Deep Learning Algorithm Predicts Future DR Severity

A new fully automated computerized detection algorithm applied to ultrawide field (UWF) images can identify the risk of diabetic retinopathy (DR) progression associated with predominantly peripheral hemorrhages and microaneurysms (PPL-HMA), according to new study results presented at the 2018 meeting of the Association for Research in Vision and Ophthalmology (ARVO).

The algorithm has the potential to allow ophthalmologists to custom tailor follow-ups individually to patients, according to Paolo Silva, MD, assistant professor of medicine at Harvard Medical School, as well as predict progression of diabetic retinopathy over 4 years.

Paolo Silva, MD:

A lot of my work is on Ultra Wide Field (UWF) retinal imaging. We've been developing this technology in telemedicine programs.

What this study looked at is if we can automate the way we evaluate UWF, specifically looking for hemorrhages and micro aneurysms and whether these hemorrhages and micro aneurysms predict progression [of diabetic retinopathy] over 4 years.

The data suggest that with a fully automated computerized algorithm for detecting hemorrhages and micro aneurysms, we're able to better predict the risk for progression over 4 years. This meant when you see hemorrhages in the retinal periphery, when these are present in a way that the H/Mas, or the hemorrhages and micro aneurysms are predominantly outside the ETDRS 7 fields, the risk for progression may be as high as 4-fold increase.

The presence of these predominantly peripheral lesions (PPL) potentially may increase, or its absence may decrease the risk for progression, so this adds to our ability to better stratify, better prognosticate patients with diabetic retinopathy.

This is a fully automated way. In eyes with diabetic retinopathy, the hemorrhages and micro aneurysms can be as little as a single micro aneurysm, but can be as numerous as 1,000, so counting these individually is completely not feasible, so an automated way is able to give us a precise number. And by localizing where these H/Mas are, they're able to help us better predict which eyes are at risk for progression and which eyes aren't.

This can custom tailor your follow ups for these patients. Patients can be followed up for longer intervals or can potentially be followed up for shorter intervals.

First, the UWF technology is very expensive. The automated algorithm is not commercially available, but what this data suggests is that this technology is available, and in the future may potentially change the way we evaluate and assess eyes with diabetic retinopathy.

With deep learning algorithms, what we can potentially do is find lesions that we have not been looking at before. If we find these lesions, these lesions can help clinicians predict progression, and also potentially a new lesion that we haven't seen before. Something in the retina that no one's been seeing, or no one's been able to see.

Now, with deep learning algorithms, we can predict progression.

Get frontline clinical insights directly to your inbox.

Related Coverage >>>

Several Gene Therapies on Horizon for Retinitis Pigmentosa

Outcomes Worse with anti-VEGF than PRP in Patients Lost to Follow-Up

Artificial Intelligence Effectively Assesses Cell Therapy Functionality