Source-linked AI summary
Intra-Retinal Layer Segmentation of 3D Optical Coherence Tomography Using Coarse Grained Diffusion Map
Raheleh Kafieh, Hossein Rabbani, Michael D. Abramoff, Milan Sonka
TL;DR
OCT retinal-layer segmentation remains difficult, especially when image contrast or layer gradients are weak. This paper uses sequential diffusion maps on 2D and 3D SD-OCT data and reports low border-positioning errors, with additional robustness across scanners and small artifacts.
Problem
OCT segmentation remains difficult, while existing approaches have limited comparability and may rely on intensity or edge information.
Method
The method applies two sequential diffusion maps: the first isolates the ILM-to-RPE region, and the second localizes internal retinal layers using regional texture.
Results
The mean unsigned border positioning errors were 8.52 - 3.13 micrometer for 2D and 7.56 - 2.95 micrometer for 3D methods.
Takeaways & Limitations
The approach outperformed several published graph-based methods and was reported as robust across scanners and against small artifacts.
Takeaways & Limitations
The study did not consider retinal layer pathologies, and graph-node size requires a tradeoff between accuracy and noise management.
Abstract
from arXiv · showhide
Optical coherence tomography (OCT) is a powerful and noninvasive method for retinal imaging. In this paper, we introduce a fast segmentation method based on a new variant of spectral graph theory named diffusion maps. The research is performed on spectral domain (SD) OCT images depicting macular and optic nerve head appearance. The presented approach does not require edge-based image information and relies on regional image texture. Consequently, the proposed method demonstrates robustness in situations of low image contrast or poor layer-to-layer image gradients. Diffusion mapping is applied to 2D and 3D OCT datasets composed of two steps, one for partitioning the data into important and less important sections, and another one for localization of internal layers.In the first step, the pixels/voxels are grouped in rectangular/cubic sets to form a graph node.The weights of a graph are calculated based on geometric distances between pixels/voxels and differences of their mean intensity.The first diffusion map clusters the data into three parts, the second of which is the area of interest. The other two sections are eliminated from the remaining calculations. In the second step, the remaining area is subjected to another diffusion map assessment and the internal layers are localized based on their textural similarities.The proposed method was tested on 23 datasets from two patient groups (glaucoma and normals). The mean unsigned border positioning errors(mean - SD) was 8.52 - 3.13 and 7.56 - 2.95 micrometer for the 2D and 3D methods, respectively.
1. INTRODUCTION
OCT segmentation is important but difficult because no single method works equally well across tasks and reported approaches use different settings. Prior work spans intensity, active-contour, pattern-recognition, and graph-based methods, with graph approaches reporting advantages in accuracy and processing time.
- Segmentation challenge: OCT segmentation remains difficult despite being a commonly required step in OCT image analysis.The paper notes that no typical segmentation method is expected to work equally well for all tasks.
- Prior approaches: Prior OCT segmentation methods include A-scan, B-scan, active-contour, pattern-recognition, and 3D graph-based approaches.These categories classify methods by the image domain subjected to segmentation.
- Prior approaches: Pattern-recognition methods reported different tradeoffs, including 6 pixels of line difference and 2 minutes for one method versus 2 pixels and 45 seconds for another.The latter method was reported to improve both linear difference and processing time relative to the former.
- Prior approaches: Graph-based approaches were reported to outperform earlier intensity, active-contour, and pattern-recognition techniques in accuracy and processing time.The paper describes 3D graph-based methods as suitable for OCT segmentation relative to the approaches discussed.
- Comparison caution: Reported method-comparison numbers cannot be directly compared because the evaluated methods used different settings.This qualification accompanies the brief comparison table of OCT segmentation approaches.
2. DIFFUSION MAPS
Diffusion maps represent graph nodes through Markov-transition geometry and spectral coordinates, enabling clustering in a Euclidean embedding. For image segmentation, the method constructs nodes from grouped pixels or voxels, combines geometric and feature distances, sparsifies and normalizes graph weights, and clusters the resulting diffusion coordinates.
- Diffusion-map representation: Diffusion maps embed graph nodes into Euclidean space using eigenvalues and eigenfunctions of a normalized Markov matrix.Clustering methods such as k-means can then be applied in this transformed space.
- Diffusion-map representation: The diffusion time τ sets an analysis scale, and this application chooses τ=1 rather than analyzing multiple scales.The paper connects diffusion-map geometry to transition probabilities over τ steps.
- Coarse graining: Coarse graining groups graph nodes into partitions while preserving similar spectral properties and averaging transition weights between groups.The partition is selected by minimizing quantization distortion in diffusion space.
- Coarse graining: Diffusion-space distortion measures weighted distances between embedded nodes and their partition centroids, providing a criterion for clustering quality.The formulation is also related to weighted pairwise distances and kernel k-means.
- Image implementation: Image nodes need not be individual pixels or voxels; aggregating groups can increase feature dimensionality while helping manage noise.The paper presents grouped pixels or voxels as graph nodes for image segmentation.
- Image implementation: Graph construction uses geometric distances between points and feature distances between their features, with separate scale factors for the two distance types.The recommended scale factor is 0.15 times the corresponding distance-matrix range.
- Image implementation: A neighborhood radius suppresses non-neighbor weights to create a sparse matrix, which is then row-normalized into a Markov matrix before eigendecomposition and clustering.The implementation applies k-means to diffusion coordinates and repeats it to minimize quantization distortion through coarse graining.
3. IMPLEMENTING DIFFUSION MAPS ON GRAY-LEVEL IMAGES
The method constructs graph nodes from grouped OCT pixels or voxels, then applies diffusion-map clustering and post-processing to localize retinal boundaries. Two sequential stages progressively narrow the search region and refine segmentation.
- Graph construction: 10 × 10 pixel boxes form graph nodes for the first 2D diffusion-map stage.The method associates graph nodes with image pixels and uses rectangular windowing before clustering.
- Sequential segmentation: The first stage segments six boundaries concurrently: the 1st and 7th–11th retinal boundaries.The second stage identifies the inner boundaries, including the additional boundary labeled 6a.
- Graph construction: Graph weights combine geometric distance between box centroids with differences in mean gray level.The scale factors for these distances are data-dependent and recommended as 0.15 times their respective ranges.
- Clustering: k-means clustering with k=3 is applied in diffusion-coordinate space, followed by coarse graining to reduce sensitivity to randomized initial centroids.Coarse graining minimizes quantization distortion in the diffusion space.
- Boundary refinement: Cluster edge points are interpolated and then shifted toward the lowest vertical gradient to compensate for box-size localization error.Subsequent boundary processing includes smoothing and local regression.
4. IMPLEMENTING DIFFUSION MAPS ON 3D OCT DATA
The 3D method applies diffusion maps to cubic OCT graph nodes in two successive stages, using coarse-grained spectral clustering to extract outer and inner retinal boundaries. Selected voxel cubes are represented by centroid geometry and mean gray-level information, then aligned and interpolated into final 3D surfaces.
- Graph construction: 3D graph nodes are cubic portions of the OCT volume, with 10 × 10 × 10 pixel cubes used in the first stage.The evaluated volumes included 200 × 200 × 1024 and 512 × 650 × 128 pixels.
- Two-stage segmentation: The first diffusion-map stage extracts six boundaries concurrently: the 1st and 7th through 11th boundaries.The second stage targets the 2nd through 6th boundaries and the additional 6a boundary.
- Graph construction: Each cube is characterized by its centroid coordinates and mean gray level, with separate geometric and gray-level scale factors.The scale factors are calculated as 0.15 times the corresponding distance-matrix ranges.
- Spectral clustering: The spectral coordinates are clustered with k-means using k=3, followed by coarse graining of the eigen-function space.The first and second diffusion maps are each visualized through their three most important eigenfunctions and cluster centroids.
- Surface formation: Detected boundaries are obtained on selected slices, interpolated across remaining slices, and aligned to form final 3D segmentation surfaces.Cubic spline smoothing, weighted local linear regression, and second-degree polynomial models are applied sequentially to 3D edge points.
5. EXPERIMENTS AND RESULTS
The method was evaluated on glaucoma and normal-eye OCT datasets using manual observer tracings and boundary-positioning errors. Results showed significant improvement over cited algorithms for nearly all compared glaucoma boundaries, while 2D and 3D methods had boundary-specific trade-offs on normal-eye data.
- Datasets: 23 OCT datasets were evaluated: 10 glaucoma datasets and 13 normal-eye macular datasets.The glaucoma volumes contained 7 manually traced retinal layers, while the normal-eye volumes were evaluated for 10 layers and 11 surfaces.
- Glaucoma test set: The glaucoma evaluation compared automated segmentation with manual tracings from two independent observers on randomly selected slices.The algorithm was tested on seven surfaces because that was the subset available in the independent standard.
- Glaucoma results: The proposed method showed significant improvement over Lee et al. and Antony for all reported comparisons except Lee et al. on the 6th boundary.Mean signed and unsigned border-positioning errors and mean layer-thickness errors were computed.
- Isfahan test set: The normal-eye evaluation used mean signed and unsigned border-positioning errors for each surface, with manual tracing by two observers as the validation reference.Examples were reported for both the 2D and 3D approaches.
- 2D–3D comparison: Gradient-based boundaries were better localized in 2D, whereas boundaries 2–6a had lower localization error with the global 3D method.The 3D procedure reduced computation by applying gradient search only to selected slices and interpolating the remaining slices.
- Missed-boundary handling: Very faint boundaries may not be distinguished, so the method can either ignore missed layers or apply a new diffusion map within the relevant bounded region.For the second strategy, information between neighboring boundaries is removed before constructing a new graph and diffusion map.
6. DISCUSSION AND CONCLUSION
The proposed diffusion-map method segments intraretinal layers in 2D and 3D OCT and remains robust across scanners and small blood-vessel artifacts. Its main scope boundary is that fixed thin graph regions may be unsuitable for pathological, non-smooth retinal layers.
- Discussion and conclusion: The method uses two sequential diffusion maps to isolate the ILM-to-RPE complex and then localize internal retinal layers.The first map reduces the data to the region of interest; the second identifies layers within it.
- Discussion and conclusion: Experiments suggested robustness across Zeiss Cirrus and Topcon 3D OCT-1000 scanners, while the global approach was intrinsically insensitive to small artifacts.Blood-vessel-artifact robustness is illustrated in Fig. 30.
- Discussion and conclusion: The graph-region size is an important parameter: thin rectangles or prisms suit smooth aligned layers but may fail when pathological surfaces are not smooth.The paper presents pathologic data as an example of this problem.
- Discussion and conclusion: Selecting smaller graph nodes improves localization accuracy but reduces noise cancellation, creating a tradeoff between accuracy and noise management.The method averages pixels or voxels within each node to suppress noise.
- Discussion and conclusion: The proposed method achieved meaningful frequency-domain clustering with a subsequent gradient search, whereas the compared Ghorbel approach required extra boundary-extraction steps.The sparse representation also keeps time complexity low by reducing image resolution.