# ReticularNet: Automated Pixel-Level Segmentation of Reticular Pseudodrusen on Near-Infrared Reflectance Images by Deep Learning

## TL;DR

This paper introduces ReticularNet, a deep learning model that automatically segments reticular pseudodrusen in near-infrared images, outperforming human experts in accuracy and consistency.

## Contribution

A novel deep learning model for automated, pixel-level segmentation of reticular pseudodrusen in clinical near-infrared images.

## Key findings

- ReticularNet achieved a mean Dice similarity coefficient of 0.36, significantly higher than human experts.
- The model showed higher consistency in lesion number, pixel area, and contour area compared to ophthalmologists.
- ReticularNet's performance was validated across a diverse dataset of 508 images from 117 eyes.

## Abstract

Reticular pseudodrusen (RPD) represent an important biomarker in age-related macular degeneration (AMD) but are difficult to grade and often assessed only for presence or absence, without quantitative or spatial analysis of RPD burden. The objective was to develop and validate a deep learning model for pixel-level RPD grading on near-infrared reflectance (NIR) images, which are commonly acquired in clinical practice and the most accurate en face detection modality.

Deep learning model development study.

Five hundred eight images of 117 eyes (70 participants) with or without RPD, over a wide range of AMD severities.

The ground truth grading pipeline comprised reading center multimodal grading for RPD presence and NIR annotation with RPD contours, followed by pixel-level NIR annotation of all individual RPD lesions. The data set was split 80:20 into training and test sets. A DeepLabv3-ResNet-18 segmentation deep learning model (“ReticularNet”) was trained to perform pixel-level grading of RPD on NIR images. Its performance was compared with that of 4 ophthalmologists.

Dice similarity coefficient (DSC); intraclass correlation coefficient (ICC) for RPD lesion number, pixel area, and contour area.

For pixel-level grading, ReticularNet achieved a mean DSC of 0.36 (standard deviation 0.16). This was significantly higher than the mean DSC of each ophthalmologist (0.03, 0.13, 0.19, and 0.23; P ≤ 0.02 for each) and of all ophthalmologists together (P < 0.0001). ReticularNet had ICCs of 0.44 (lesion number), 0.56 (pixel area), and 0.61 (contour area), with no significant underestimation or overestimation (P ≥ 0.24). These values were numerically higher than the ICCs of each ophthalmologist, who had ICC ranges of –0.08 to 0.23, –0.05 to 0.40, and –0.09 to 0.58, respectively, and significant underestimation in almost all cases. For all 3 parameters, ReticularNet’s ICC was significantly higher than that of all specialists considered together (P ≤ 0.02).

ReticularNet achieved automated pixel-level grading of RPD on NIR images. Its grading was superior to that of 4 ophthalmologists, across a variety of metrics. We are making the code/models available for research use. Improved access to quantitative and spatial RPD grading should lead to improved understanding of these lesions as important biomarkers of retinal disease.

Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.