Abstract
Purpose :
With the advent of high-resolution optical coherence tomography (OCT) devices, retinal layers previously non-distinguishable are becoming identifiable. The challenge we address is how to efficiently adapt retinal layer boundary segmentation algorithm to high resolution and to additionally regress novel retinal boundaries as part of their output.
Methods :
Our retinal layer segmentation method is based on SD-LayerNet, a deep learning-based architecture, which imposes topological constraints on the layer boundary positions, allowing a restriction on the possible location of the novel layer to be regressed. In addition, we introduce a resolution consistency constraint, which assures that the predictions on the high-resolution scan are the same as on the standard resolution one. This is utilized to transfer the layer segmentation knowledge between imaging devices of different resolution. We illustrate the paradigm on the simulated case of segmenting a weakly distinguishable boundary between ganglion cell layer (GCL) and inner plexiform layer (IPL). Namely, the method is trained on a dataset acquired with Spectralis OCT with a set of layer annotations that do not include GCL/IPL boundary and is fine-tuned in a few-shot manner to the dataset acquired with High-Res OCT (Heidelberg Engineering), where GCL/IPL is annotated.
Results :
A total of 50 OCT volumes acquired with the Spectralis OCT device were used as a development set. 34 High-Res OCT of patients with age-related macular degeneration (AMD) were manually labeled including the GCL/OPL retinal boundary. Five High-Res scans were used for fine-tuning and 29 for the final validation. In the simulation of a novel boundary (GCL/IPL), the algorithm achieved a mean absolute error (MAE) of 6.26 ± 2.61 µm. This was comparable to the lower error bound of 5.33 ± 2.39 µm by the method trained on a larger High-Res dataset in a fully supervised manner (Figure 1).
Conclusions :
Our SD-LayerNet demonstrated very efficient fine-tuning properties for regressing potentially novel retinal layers on High-Res OCT. This is an important step toward expanding the repository of retinal layers that can be automatically segmented with the introduction of high-resolution OCT devices.
This abstract was presented at the 2024 ARVO Annual Meeting, held in Seattle, WA, May 5-9, 2024.