Deep Learning-based Multi-Class Fluid Segmentation in Swept-Source OCT Images

We have now submitted our recent work on the deep learning-based segmentation of intra-retinal (IRF), sub-retinal (SRF) and serous pigment epithelial detachments (PEDs) in swept-source optical coherence tomography (SS-OCT) images.  To the best of our knowledge, we are the first to publish such an algorithm using this increasingly important modality.  In this work we show that this fully automated approach segments these fluid types in a heterogeneous wet age-related macular degeneration (wAMD) population with accuracies matching expert graders.  Combined segmentations of IRF and SRF with sub-retinal pigment epithelium (RPE) fluid in serous PEDs, offer a near complete automated solution for retinal fluid in wAMD.  An automated approach to multi-layer fluid segmentation offers an advantage to physicians managing patient care amidst a busy clinical setting, particularly with the growing disease burden of wAMD. Furthermore, applying this deep-learning approach in clinical trials for therapeutic drug development could accelerate trial end points and the development of novel therapeutics within commercial timelines.

A pre-print of this work is available here.

figure labels 02 Sep 2020 09 27 51
Example IRF (red), SRF (blue) and PED (green) results: input images (A,B,C,D); ground truth labels (E,F,G,H); automated segmentations (I,J,K,L).