Automated Deep Learning-based Multi-class Fluid Segmentation in Swept-Source Optical Coherence Tomography Images

Abstract

PurposeTo evaluate the performance of a deep learning-based, fully automated, multi-class, macular fluid segmentation algorithm relative to expert annotations in a heterogeneous population of confirmed wet age-related macular degeneration (wAMD) subjects.MethodsTwenty-two swept-source optical coherence tomography (SS-OCT) volumes of the macula from 22 from different individuals with wAMD were manually annotated by two expert graders. These results were compared using cross-validation (CV) to automated segmentations using a deep learning-based algorithm encoding spatial information about retinal tissue as an additional input to the network. The algorithm detects and delineates fluid regions in the OCT data, differentiating between intra- and sub-retinal fluid (IRF, SRF), as well as fluid resulting from in serous pigment epithelial detachments (PED). Standard metrics for fluid detection and quantification were used to evaluate performance.ResultsThe per slice receiver operating characteristic (ROC) area under the curves (AUCs) for each of these fluid types were 0.90, 0.94 and 0.94 for IRF, SRF and PED, respectively. Per volume results were 0.94 and 0.88 for IRF and PED (SRF being present in all cases). The correlation of fluid volume between the expert graders and the algorithm were 0.99 for IRF, 0.99 for SRF and 0.82 for PED.ConclusionsAutomated, deep learning-based segmentation is able to accurately detect and quantify different macular fluid types in SS-OCT data on par with expert graders.

Validation and Clinical Applicability of Whole-Volume Automated Segmentation of Optical Coherence Tomography in Retinal Disease Using Deep Learning

“…27,28 Other reasons for higher DSCs in related work may be owing to a lack of independence of scans between the development and evaluation set, 29 and the derivation of ground truth labels from the model being evaluated. 30 Our study used a pool of 6 graders, all blinded to model predictions, with 2 graders segmenting each OCT volume. This provided multiple clinical interpretations resulting in the ability to benchmark the model against the natural, rich variation seen in clinical practice.…”

“…In related work, models developed and evaluated on single 2-D sections have achieved higher DSCs, where exhaustive grading can be performed more feasibly . Other reasons for higher DSCs in related work may be owing to a lack of independence of scans between the development and evaluation set, and the derivation of ground truth labels from the model being evaluated . Our study used a pool of 6 graders, all blinded to model predictions, with 2 graders segmenting each OCT volume.…”

Utilization of deep learning to quantify fluid volume of neovascular age-related macular degeneration patients based on swept-source OCT imaging: The ONTARIO study

“…Given these were large input volumes and a heterogeneous input data set, the reported validation compared excellently to manual segmentations. Results showed strong correlation in all fluid volumes between the algorithm and the manually labeled data [ 17 ].…”

“…A deep learning-based algorithm was implemented to segment fluid regions within each OCT-volume. As reported previously, the method takes each OCT B-scan as input, and a segmentation mask is generated automatically using Orion software (San Francisco, CA, USA) [ 16 , 17 ]. The deep learning architecture used for first the training, and subsequently the testing, was that of U-Net, which is a version of the autoencoder that uses skip connections to better maintain detail across different scales [ 18 , 19 ].…”

Quantitative Assessment of Automated Optical Coherence Tomography Image Analysis Using a Home-Based Device for Self-Monitoring Neovascular Age-Related Macular Degeneration

“…Layer segmentation was performed the existing Orion software, previously described, 11 and the results were automatically exported as binary masks for comparisons using the Dice similarity coefficient (DSC), which measures the degree of agreement between the automated segmentation (A) and the ground truth (B), based on the area of overlap: DSC ¼ 2jA\Bj jAjþjBj . These automatically generated binary masks were additionally used as input to the deep learning approach previously validated for use with swept source OCT 12 to segment the fluid regions. This U-Net like architecture 13 -an autoencoder with skip connections -takes as input both the OCT B-scans and a retinal layer segmentation mask from Orion and has been successfully applied to similar image segmentation tasks.…”