Prediction of retinopathy progression using deep learning on retinal images within the Scottish screening programme

Background/Aims
National guidelines of many countries set screening intervals for diabetic retinopathy (DR) based on grading of the last screening retinal images. We explore the potential of deep learning (DL) on images to predict progression to referable DR beyond DR grading, and the potential impact on assigned screening intervals, within the Scottish screening programme.

Methods
We consider 21346 and 247233 people with T1DM and T2DM respectively each contributing on average 4.8 and 4.4 screening intervals of which 1339 and 4675 intervals concluded with a referable screening episode. Information extracted from fundus images using DL were used to predict referable status at the end of interval and its predictive value in comparison to screening-assigned DR grade was assessed.

Results
The DL predictor increased the AUC in comparison to a predictor using current DR grades from 0.809 to 0.87 for T1DM and from 0.825 to 0.87 for T2DM. Expected sojourn time – the time from becoming referable to being rescreened - was found to be 3.4 (T1DM) and 2.7 (T2DM) weeks less for a DL-derived policy compared to the current recall policy.

Conclusions
We showed that, compared to using the current retinopathy grade, DL of fundus images significantly improves the prediction of incident referable retinopathy before the next screening episode. This can impact screening recall interval policy positively, for example, by reducing the expected time with referable disease for a fixed workload - which we show as an exemplar. Additionally, it could be used to optimise workload for a fixed sojourn time.