Comparison of procedural and neural network algorithms for segmentation of regions of non-perfusion in retinal OCTA scans

Abstract

Purpose : Segmentation of regions of nonperfusion (RNP) with an automated algorithm enables detection of these regions by personnel that have not been trained in interpreting OCTA scans.
The procedural, or conventional, algorithm uses vessel binarization and image processing to detect areas without vessels. This algorithm is subject to error in regions with noise or low signal. The alternative algorithm is a convolutional neural network (CNN) that has been trained with OCTA retina projections and OCT reflectance images as inputs and annotations made by expert clinicians as targets.

Methods : The CNN was trained using 60,000 patches obtained from 14 15x15 and 46 12x12 OCTA scans (PLEX® Elite 9000, ZEISS, Dublin, CA) of eyes having significant areas of ischemia, along with expert segmentations of RNPs. 21 scans not used for the training were segmented using the procedural and DL algorithms. Outputs were compared against annotations of these scans made by expert clinicians. Dice coefficients were calculated for each output vs the expert segmentation for each acquisition.

Results : For the 21 scans in the test dataset, the CNN outperformed the procedural algorithm. Undesired segmentation of the FAZ as RNP was reduced but not eliminated in the CNN output, and some regions of low signal were incorrectly segmented. This may be improved by modifying the CNN training dataset and architecture.

Conclusions : A CNN algorithm for segmentation of RNP compares favorably with one based on binarization and thresholding of OCTA scans.

This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.

 

Dice coefficients between the algorithm outputs and the expert segmentation are compared in this Bland-Altman plot. The vertical axis is the difference between the Dice coeffs obtained with the CNN and those from the procedural algorithm. All are positive, showing that the CNN produced better agreement than the procedural algorithm.

 

Example scan of an eye exhibiting non-perfusion. Left: Retina slab without segmentations. Center: comparison of procedural algorithm and expert annotation. Green: areas segmented by both the algorithm and the expert. Blue: segmented by the expert but not the algorithm. Red: segmented by the algorithm and not the expert. Right: comparison of CNN output and the expert. CNN shows greater agreement and a higher Dice coefficient (0.81 vs 0.70).