Performance of a novel deep learning algorithm for Automatic Retinal Fluid Quantification in Home OCT Images

Abstract

Purpose : To evaluate the performance of a deep learning algorithm in identifying and quantifying retinal fluid in output of a patient self-operated Notal Home OCT (NOCT) device

Methods : Data were collected from 5 clinics where AMD patients self-imaged the central 10° of their maculae with the NOCT V2.5. 355 eyes from 239 subjects were enrolled, mean age 78 years (range: 54-92). From each cube scan (88 B-scans), avg. of 10 B-scans were selected for manual segmentation. Each B-scan was labeled pixel-wise into 4 compartments: Vitreous/outer layers (V/O), Retina (R), subretinal fluid (SRF), intraretinal fluid (IRF). Eyes were randomly split into learning and validation sets (ratio 8/1). Quantifier was developed using semantic segmentation with convolutional neural network. 1)Fluid quantification was evaluated by correlating each B-scan’s fluid area segmented by human vs machine. 2)Pixel-wise fluid was compared withrecall / precision. 3)Presence of fluid in B-scans were compared for accuracy.

Results : 3428 B-scans were manually segmented (75% with fluid). The learning set included 2936 B-scans of 311 eyes.The validation set included 492 B-scans of 44 eyes.
The area of SRF and IRF from the automatic segmentation were compared to the human-defined segmented area: 1) Pearson correlation of fluid area was 0.98 (p<0.00001) for SRF and 0.90 for IRF (p<0.00001). 2) SRF Fluid pixel-wise recall was 0.72 and precision was 0.86; The IRF recall was 0.80 and the precision was 0.77. 3) Sensitivity for detecting presence of SRF in B-scans was 0.985 and specificity 0.98. Sensitivity for detecting IRF was 0.99 and specificity was 0.97.

Conclusions : This version of NOA performed comparably to human readers, utilizing deep learning for automatic fluid segmentation and quantification on self-obtained NOCT V2.5 images, as part of a home-based OCT fluid-evaluation system. By analyzing fluid status and its dynamics, and visualizing the locations of fluid, this system has the potential to accurately monitor AMD disease activity in a patient self-operated home-use environment.

This is a 2020 ARVO Annual Meeting abstract.

 

Figure 1, SegmentationResults: (a) Two B-scans from the validation set. (b,c) Manual and auto-segmentation of the B-scans. (d) Fluid thickness maps -fluid distribution for SRF and IRF extracted from the auto-segmentation. Blue arrow indicates the position of the displayed B-scans (a) in the map. Green circles with diameter of 1 and 3 mm.