Autoencoders are another type of DL that can enhance the analysis of OCT images through advanced dimensionality reduction and feature extraction. These neural network architectures compress complex OCT data into a lower-dimensional space, isolating critical features that improve diagnostic accuracy and disease progression assessment. In a study by Shon et al., variational autoencoders were used to analyze anterior segment OCT images, uncovering hidden patterns and features that enhance the ability to differentiate between various stages of glaucoma and improve diagnostic precision.
33 The model analyzed data from 2111 eyes, successfully identifying latent variables that captured features such as anterior chamber area, corneal curvature, and pupil size. One of these variables, representing complex interactions among multiple anterior segment structures, was significantly smaller in PACG eyes compared with PAC eyes (
P = 0.015), highlighting its potential for detecting subtle differences in disease progression. Similarly, Bowd et al. utilized an autoencoder to extract and compress features from OCT images to enhance the classification and improve the assessment of glaucomatous changes.
34 The study demonstrated that the DL-based regions of interest (ROIs) identified progression in glaucoma eyes with a sensitivity of 90%, significantly higher than the 63% sensitivity achieved using global RNFL thickness measurements. Additionally, the ROIs showed faster rates of structural change in progressing eyes compared to conventional measurements (−1.28 µm/year vs. −0.83 µm/year), emphasizing the method's potential for earlier and more precise detection of disease progression.
34 Furthermore, Panda et al. demonstrated the use of autoencoder networks in SD-OCT image analysis, effectively reducing dimensionality and enhancing image quality for improved disease detection and characterization of novel biomarkers for glaucoma.
35 The study achieved a diagnostic accuracy of 92% with a sensitivity of 90% at 95% specificity, demonstrating the network's robust performance. By altering principal components in the latent space, the researchers illustrated how ONH morphology shifts from a “nonglaucoma” to a “glaucoma” condition, linking these structural changes to clinical observations and identifying novel biomarkers for glaucoma diagnosis.
35 By concentrating on the most relevant features, these studies illustrate how autoencoders can improve diagnostic precision. Additionally, autoencoders can reduce the required training set size and computing resources, making OCT analysis more accessible and efficient. Integrating autoencoders with other DL models could offer a comprehensive approach to OCT image analysis, providing detailed insights into the structural characteristics associated with glaucoma and other ophthalmic conditions.