Automatic segmentation and classification of intraretinal cystoid fluid and subretinal fluid in 3D-OCT using convolutional neural networks

Abstract

 

Purpose

 

The key driver for vision loss in macular diseases such as neovascular age-related macular degeneration is the accumulation of macular edema. Out of the two components of macular edema, intraretinal cystoid fluid (IRC) leads to severe vision loss. Conversely, recent evidence suggests that subretinal fluid (SRF) may be associated with better visual acuity outcomes. The precise classification and quantification of IRC and SRF are therefore of paramount importance for disease management. We propose a fully automated segmentation and classification method for IRC and SRF in spectral-domain optical coherence tomography (SD-OCT) images.

 

Methods

 

We use convolutional neural networks (CNN) to capture characteristic visual appearance patterns and classify normal retinal tissue, IRC and SRF. The CNN is trained in a supervised manner using approximately 300,000 2D image patches extracted from 157 OCT image volumes available at the Vienna Reading Center. All image patches are sampled at random positions. 73% of the patches show healthy tissue, 8% show IRC and 18% show SRF. Besides the visual appearance of the image patches we provide the CNN with the Euclidean distances of the patch centers to the fovea and the 3D coordinates of the patch centers. We evaluate the automatic segmentation performing four-fold cross-validation. Based on the image patches of the training set the CNN learns representative and discriminative features appropriate for pixel-wise image classification. The pixel-wise classification results in a segmentation of the whole OCT volume into normal retinal tissue, IRC and SRF.

 

Results

 

The amount of spatial overlap between predicted class labels and the corresponding ground truth annotations is computed as performance measure. The classifier achieves an overall accuracy over all three classes of 96%. The pixel-based class-wise accuracies of normal retinal tissue, IRC and SRF are 98%, 90% and 92%, respectively.

 

Conclusions

 

Our CNN classifier automatically and highly accurately segments and discriminates between normal retinal tissue, IRC and SRF in retinal SD-OCT. This may enable precise structure-function correlations, and the prediction of visual function based on SD-OCT on the large scale.  

 

Figure 1: Intensity image (a), ground truth (b) and segmentation result (c) of a case showing SRF (top) and IRC (bottom).