Fully Automated Analysis of Retinal Images by Deep Learning Holds Promise
Testing their system against expert manual analysis of OTC scans, the researchers determined that the automatic diagnostic method was both reliable and accurate.
Thomas Schlegl, MSc
Interest in the promise of deep learning to assist in the analysis of optical coherence tomography (OCT) scans led Thomas Schlegl, MSc, with the Christian Doppler Laboratory for Ophthalmic Image Analysis at the Medical University Vienna, in Austria, and colleagues to develop and validate a fully automated method of detecting and quantifying macular fluid in OTC imaging.
The research group developed an accurate automated method of detecting and quantifying intraregional cystoid fluid (IRC) and subretinal fluid (SRF) in 3 macular pathologies: neovascular age-related macular degeneration (AMD), diabetic macular edema (DME), and retinal vein occlusion (RVO). Testing their system against expert manual analysis of OTC scans, the researchers determined that the automatic diagnostic method was both reliable and accurate, providing a "promising horizon" for clinical ophthalmology diagnosis and treatment.
Schlegl and colleagues pointed out that the use of OCT as both a diagnostic and guiding therapy tool for macular pathologies like AMD, DME, and RVO, has "profoundly disrupted" the ophthalmological field's understanding of macular disease. Clinical use of OCT has increased rapidly due to the benefits of advanced OCT technology in diagnosis and treatment, but as a result of increased use "the feasibility of manual OCT assessment in clinical practice has become largely unrealistic."
They stated that there is an unmet need for a reliable, automated means of analyzing OCT images “beyond a purely anatomic presence/absence detection.”
What is needed, according to Schlegl, is an accurate and reliable means of automating the measurement of markers for disease activity, such as the presence of IRC and SRF through OCT images, which can help clinicians in establishing individualized therapy regimens for patients, and serve as prognostic markers for visual acuity (VA) in patients.
The automated system Schlegl and colleagues developed uses convolutional neural networks, sometimes referred to as deep-learning or machine learning algorithms, to perform analyses of IRC, SRF, and "non-fluid regions" in a standard OCT image.
The software uses 2 processing components using an "encoder-decoder" architecture, according to Schlegl. It "encodes," or creates an abstract representation of an OCT image, and then "decodes," or maps that abstract image in comparison to clinical class labels. The system was trained to determine normal tissue, IRF, or SRF through embedded data, and then predict the correct clinical class label for every pixel in the OCT image.
Using a clinical dataset of 1200 OCT volumes of patients with AMD (n = 400), DME (n = 400), or RVO (n = 400), the researchers tested the ability of the automated system to detect and quantify IRC and SRF against a manual reading of OTC scans.
In order to test the accuracy of performance, "we computed receiver operating characteristic (ROC) curves by varying the threshold over the number of fluid pixel segmented by our model," Schlegl stated.
Using the ROC curves, Schlegl and colleagues designated the area under the curve (AUC) as the mean sensitivity value ranges from 0.50 (discriminative performance equal to chance) to 1.00 (perfect discriminative performance).
The study determined that the automated diagnostic method achieved optimal accuracy for IRC in all 3 macular pathologies with a mean accuracy of 0.94, a mean precision of 0.91, and a mean recall of 0.84. The automatic diagnostic method achieved high accuracy for SRF for AMD and RVO patients with a mean accuracy of 0.92, a mean precision of 0.61, and a mean recall of 0.81.
Because DME was represented rarely in the population studied, Schlegl reported that detection of DME showed poorer overall performance in comparison to AMD and RVO.
The high level of concordance between the automated diagnostic method and manual assessment of OTC images by experts suggests that "fully automated analysis of retinal OCT images from clinical routine provides a promising horizon" for the future of clinical ophthalmology, according to the authors.
As Schlegl writes, "the management of the leading exudative macular diseases by vascular endothelial growth factor (VEGF) inhibition is largely based on the evaluation of retinal fluid for initial diagnosis and retreatment indications," making OCT imaging a "mainstay" in clinical settings. However, the proliferation of OCT imaging use in busy clinics has made it "inherently impractical" to rely on manual image inspection and analysis.
The system that Schlegl and colleagues have created may help lift some of the burdens of diagnosis and analysis of OCT images by quickly and accurately detecting fluid in the majority of those images and directing the attention of clinicians to specific images that require detailed expert analysis. They believe that with further testing to confirm their findings, their fully automated detection, and quantification system could become an indispensable tool for clinicians.
The study, "Fully Automated Detection and Quantification of Macular Fluid in OCT Using Deep Learning," appeared in the American Academy of Ophthalmology.
Related Coverage >>>
Treat-and-Extend Ranibizumab Regimen for AMD Comparable to Monthly Regimens
