Don't Mind the Gaps: Implicit Neural Representations for Resolution-Agnostic Retinal OCT Analysis

TL;DR

This work targets the challenge of highly anisotropic retinal OCT data by leveraging implicit neural representations (INRs) to achieve resolution-agnostic 3D analysis. It introduces two INR-based frameworks: (i) inter-B-scan interpolation guided by en-face modalities to densify sparse OCT stacks, and (ii) a population-trained, resolution-agnostic retinal atlas produced by a generalizable INR coupled with a deformation model. The proposed methods demonstrate improved retinal shape reconstruction, vessel localization, and segmentation consistency across unseen cases, outperforming linear, registration-based, and instance-based INR baselines in both healthy and CSCR datasets. By enabling continuous, modality-fused representations and atlas-based analysis, the approach has potential to enhance volumetric OCT analyses, cross-protocol applicability, and downstream tasks like registration and pathology assessment in clinical workflows.

Abstract

Routine clinical imaging of the retina using optical coherence tomography (OCT) is performed with large slice spacing, resulting in highly anisotropic images and a sparsely scanned retina. Most learning-based methods circumvent the problems arising from the anisotropy by using 2D approaches rather than performing volumetric analyses. These approaches inherently bear the risk of generating inconsistent results for neighboring B-scans. For example, 2D retinal layer segmentations can have irregular surfaces in 3D. Furthermore, the typically used convolutional neural networks are bound to the resolution of the training data, which prevents their usage for images acquired with a different imaging protocol. Implicit neural representations (INRs) have recently emerged as a tool to store voxelized data as a continuous representation. Using coordinates as input, INRs are resolution-agnostic, which allows them to be applied to anisotropic data. In this paper, we propose two frameworks that make use of this characteristic of INRs for dense 3D analyses of retinal OCT volumes. 1) We perform inter-B-scan interpolation by incorporating additional information from en-face modalities, that help retain relevant structures between B-scans. 2) We create a resolution-agnostic retinal atlas that enables general analysis without strict requirements for the data. Both methods leverage generalizable INRs, improving retinal shape representation through population-based training and allowing predictions for unseen cases. Our resolution-independent frameworks facilitate the analysis of OCT images with large B-scan distances, opening up possibilities for the volumetric evaluation of retinal structures and pathologies.

Figures (11)

• Figure 1: Top left: Interrelation between en-face image (here: SLO) and OCT. The position marked in the SLO ( ) corresponds to an A-scan (dotted line) in the OCT B-scan. Bottom left: Difference in resolution between an en-face image (here: FAF) and an analogous lateral slice of OCT. The limited amount of B-scans acquired in clinical practice leads to high anisotropy. Right: By leveraging a generalizable INR, we create resolution-independent representations of singular data instances or an atlas, respectively.
• Figure 2: Schematic overview of the proposed interpolation framework. The input for the INR are 3D coordinates in the OCT ( ) and intensities at the corresponding SLO position ( ). The INR is trained to reconstruct the intensity and segmentation label of the OCT volume at the respective coordinate. The reconstructed OCT can be evaluated at any arbitrary coordinate with a known SLO intensity. To reconstruct and segment an unseen case, the INR is kept frozen, and only the latent prior is adapted.
• Figure 3: Schematic overview of the proposed atlas generation framework. The generalizable INR predicts a deformation field based on 3D coordinates ( ) and an instance-specific latent prior. The INR layers are modulated through scale and shift vectors predicted from the latent prior by a hypernetwork (purple layers). The deformed coordinates are then input into another INR that represents the atlas with intensities and segmentation labels. The learned atlas is resolution-independent and can be evaluated on any arbitrary number of coordinates.
• Figure 4: Four examples of interpolated B-scans using linear, registration-based ehrhardt2007structure, and INR-based interpolation. The input slices for the different methods are shown with pink frames, while the original intermediate and interpolated B-scans are shown with turquoise frames. On the left side, two examples from the foveal region of the retina are depicted, showing interpolation artifacts for instance-based methods, while our GenINR manages to correctly continue the shape of the retina. On the right side, examples from the outer border of the retina are shown. Here, blood vessels are correctly localized by the GenINR with SLO integration (turquoise arrows). For the other methods, vessels can only be propagated from the input slices, again leading to interpolation artifacts.
• Figure 5: Four examples of interpolated B-scans from a CSCR patient using the proposed GenINR with FAF integration. Subretinal fluid (examples above and below left) is common in CSCR, whereas intraretinal fluid (above right) is seldom). The example below right shows a case with photoreceptor atrophy. In all cases, the retinal shape is reproduced well, but the images appear blurred in the areas of displaced retinal layers and no blood vessel shadows are visible. The color coding is the same as in Fig. \ref{['fig:qualitative_results_slo']}.