ARVO LIVE: Convolutional neural network and image quality assessment

Ophthalmology Times® talked with Mary Durbin, PhD, about a convolutional neural network that provides image quality assessment at this year's ARVO meeting.

Video transcript

Editor’s note: Transcript lightly edited for clarity.

Mary Durbin, PhD:

So I'm Mary Durbin, I'm the Vice President of Clinical Science at Topcon. I'm here at ARVO 2023, in New Orleans, to present our poster, which is titled "A Convolutional Neural Network Image Quality Assessment Applied To Screening Data."

So what we did was, as part of the Topcon Screen program, which places NW400 cameras into primary care settings in order to screen for diabetic retinopathy. The data is then read by a reading center. And this program has been very successful. So we had a large set of data that has been looked at by a reading center for diseases such as diabetic retinopathy, also other diseases. But more importantly, for our purposes, they look at the image quality. So we had an image quality assessment. And then in addition to this, we have a convolutional neural network algorithm that was designed to evaluate the image quality.

And we basically took this algorithm, which has never seen any of this data before, we applied it to the data to find out how well it would perform and by that measure how well it generalizes to a new population. And what we found was that it really performed very well. It had a sensitivity of 0.9, it had a specificity of 0.97 and an overall accuracy of 0.97. So this, you know, this algorithm performed well on unseen data, which means that it generalized well.

Not all algorithms do generalize well. But in this case, we believe it is possible to use a machine learning developed algorithm to improve an already successful program by basically telling the operator whether they need to retake an image or not retake an image, and that way they can have better workflow and more confidence in the work that they're doing, which we think is very important.