Multi-resolution Deep Learning Convolutional Networks for Improvements in OCT Retinal Layer Segmentation

Abstract

Purpose : Accurately segmenting OCT images requires both a high-resolution understanding of the specific boundary between medically significant regions, as well as an understanding of the broader context of the imaging boundary. Our group has developed an approach which uses a combination of low-magnification (low-mag) convolutional models to detect regions of interest where the specific retinal boundary is likely to be located, as well as high-magnification (high-mag) models which create high-resolution annotations within those regions of interest. The goal of this project was to measure segmentation performance improvements resulting from multi-scale approaches in images that require both large-scale contextual information and precision.

Methods : A total of 111,000 annotated OCT images were used in training and testing. For this analysis, a single retinal layer, the internal limiting membrane (ILM), was utilized for evaluation purposes. Three training methodologies were tested. One method processed OCT images using a high-resolution patch system and divided images into 128x128 patches. A second method utilized a low-mag model where training images were compressed into 128x128 patches. A third high-mag model was trained leveraging the low-mag model to center a 128x128 patch around where the low-mag model indicated the location of the retinal boundary of interest. The performance of these models was compared by measuring the average pixel distance away from manual ground truth.

Results : The patch-diced model had an average offset of 12 pixels (stdev 15 pixels) which demonstrated high resolution but poor regional context. The low-mag model had an average offset of 4.9 pixels (stdev 0.92 pixels), and the highmag-lowmag combination had an average offset of 2.6 pixels (stdev 0.93 pixels). The differences between each of the populations average distance from ground truth were statistically significant (p values less than 0.025).

Conclusions : These results demonstrate the potential for enhanced retinal layer segmentation performance through a multi-model approach using both low and high-mag system. The combined model approach provides assessment of both regional contextual information and subsequent high-resolution modeling for improved segmentation precision.

This is a 2021 ARVO Annual Meeting abstract.