Patch-based and fully semantic deep learning methods for automatic choroidal segmentation in OCT images

Abstract

Purpose : Choroidal thickness extracted from optical coherence tomography (OCT) images represents a fundamental quantitative clinical and research metric. Thus, reliable automatic methods for the segmentation of choroidal tissue represent an essential image analysis task.

Methods : A range of patch-based and fully semantic deep learning methods were developed to estimate the probability of the choroidal region of interest being present in a specific position within OCT images. A range of variables, including architecture, patch-size and image contrast enhancement, were tested to optimise segmentation performance. The network was trained using 300 images and evaluated in 294 images (from a healthy paediatric population). Manual boundary segmentation was used as a ground-truth to both train the network and evaluate the performance. The Dice overlap percentage was calculated for the choroid and retina for comparison purposes. Results were also compared with an automated standard image analysis technique.

Results : Retinal dice overlap showed a similar performance for both patch-based (99.3%±0.04%) and semantic methods (99.3%±0.01%), with marginal improvement over standard image analysis (98.8%). For choroidal boundary segmentation, semantic methods (98.0±0.05%) showed a superior performance over patch-based (97.2±0.2%), however both methods displayed significantly better performance than standard image analysis (95.9%). Changes in network architecture and contrast enhancement had minimal effect on the performance of the semantic network.

Conclusions : The segmentation performance demonstrates the potential of deep learning methods for OCT segmentation of the chorio-retinal regions, which show superior performance over standard image analysis techniques. Semantic deep learning techniques may be more suitable to the complex task of choroidal segmentation.

This abstract was presented at the 2019 ARVO Annual Meeting, held in Vancouver, Canada, April 28 - May 2, 2019.

 

Dice percentage for the two regions of interest. Each box encompasses the range of mean dice errors for all tested methods per network. Black line indicates the automatic baseline method.

 

Original B-scan images (left) along with the results provided by the semantic segmentation (right). White lines indicate ground truth.