Automated cone identification in adaptive optics retinal images of Choroideremia using a convolutional neural network

Abstract

Purpose : Adaptive optics (AO) retinal imaging has enabled noninvasive observation and quantification of the photoreceptor mosaic both in health and disease. We have previously shown that the reliability of manual cone identifications is reduced in patients with retinal disease compared to normal sighted controls. Thus, there remains a need for accurate, automated cone identification in AO images of retinal disease. Here, we assess the use of a convolutional neural network (CNN) for automatically detecting cone locations in AO images from patients with Choroideremia (CHM), an X-linked retinal degeneration.

Methods : Split-detection AO image sequences of the photoreceptor mosaic from 17 CHM patients were acquired at retinal locations surrounding the fovea and along all four meridians using an AO scanning light ophthalmoscope. Images were montaged, regions of interest (ROIs) were cropped, and cones were manually identified in each ROI using custom software. Cones in ROIs were also identified automatically using an open source CNN pre-trained on a normative population (‘normal-trained’) (Cunefare et al, Scientific Reports, 2017). True positive and false positive rates for cone identifications were determined and Dice’s coefficient was calculated for each ROI using the manual identifications as ground truth. Using a leave-one-subject-out approach, the CNN was then retrained and validated with the CHM images and manual identifications (‘CHM-trained’). Again, true positive, false positive, and Dice’s coefficient were found for each ROI in comparison to the manual identifications.

Results : The CHM-trained CNN cone identifications had a higher mean Dice coefficient than the normal-trained CNN (mean±stdev, CHM-trained: 0.821±0.100; normal-trained: 0.711±0.214). The mean true positive rate was higher for the CHM-trained CNN identifications (0.884±0.141; normal-trained: 0.642±0.260). However, the mean false discovery rate was also higher (CHM-trained: 0.201±0.128; normal-trained: 0.082± 0.075), particularly at higher eccentricities (>0.7mm) where rods were visible.

Conclusions : CNNs hold promise for automated detection of cone locations within AO images. However, CNN performance is tied closely to its training data, and performs best when operating on images in a similar domain. Our results also suggest that CNN performance will improve by training the network to consider retinal eccentricity.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.