Low-rank MERA Decomposition of 3D OCT Image Volume for Compression, Denoising, Feature Extraction and Diagnosis of Retinal Diseases

I. INTRODUCTION

• High-resolution, large-field OCT volumes burden clinical storage and transmission, while speckle noise obscures retinal detail and makes layer boundaries harder to delineate.Rare-disease datasets may also have few labeled examples, limiting deep-learning generalization,.Source: §I · pp. 1–2
• The paper asks whether a structure-aware, model-driven framework can process volumes independently and support compression, denoising, feature extraction, and diagnosis under limited supervision.This question links the imaging and data-scarcity challenges described in the introduction.Source: §I · p. 2
• The proposed CR-constrained MERA jointly supports compression, denoising, and multi-scale feature extraction while processing each volume independently without training.The approximation rank is controlled by the target compression ratio, trading compression efficiency against reconstruction fidelity.Source: §I · p. 2
• Across two scanners, low-rank MERA yields lower retinal-layer segmentation errors than noisy images and several denoising methods at compression ratios as high as 40.The paper also reports improved visual quality under substantial compression.Source: §I · p. 2
• Expert image quality metrics remain competitive with state-of-the-art denoising methods even at substantially higher compression ratios.The authors describe the reconstructions as suitable for clinical archiving, teleophthalmology, and visual assessment.Source: §I · p. 2
• On two clinical OCT datasets,, the MERA-feature GM-NSC outperforms tuned SVM and neural-network classifiers, with up to three orders of magnitude lower training complexity.Its classification performance is also reported as comparable to deep-network baselines.Source: §I · p. 2

II. BACKGROUND AND RELATED WORK

• Prior OCT denoising spans acquisition-time averaging and post-capture software methods, including filtering, sparse coding, low-rank models, and Tucker, TT, and TR tensor models.The review notes detail-preservation or computational concerns for some alternatives and motivates applying MERA to OCT volumes.Source: §II · p. 3
• Supervised deep denoising requires noisy-clean image pairs, whose acquisition is non-trivial, motivating a training-free approach for individual volumes and a classifier suited to limited supervision.The review notes involuntary eye movements and the time needed to obtain multiple scans as obstacles to collecting clean B-scans.Source: §II · p. 3

A. Notations

• Mode-Σ unfolding reshapes a tensor so modes in Σ index rows and the remaining modes index columns, preserving lexicographic order.Tensor contraction sums over shared indices to produce a tensor over the uncontracted modes.Source: A. Notations · p. 3

B. MERA Decomposition Model

• MERA is built from isometries, disentanglers, and layers formed by contracting isometries and disentanglers within the same layer.An isometry has a semi-orthonormal unfolding, whereas a disentangler has a square orthogonal unfolding.Source: B. MERA Decomposition Model · p. 3
• A MERA representation recursively combines layers and a top core; the one-layer form is denoted by the top core and its layer factors.The supplied figure illustrates the decomposition for a fifth-order tensor.Source: B. MERA Decomposition Model · pp. 3–4, Figure 1 · p. 4

C. Low-Rank MERA Approximation

• For a fifth-order tensor, the low-rank MERA approximation seeks factors at predefined ranks R1=R2=R by minimizing the Frobenius reconstruction error.The presented algorithm alternates updates of the disentangler, two isometries, and top core.Source: C. Low-Rank MERA Approximation · p. 4
• The disentangler update is an orthogonal Procrustes problem, with its solution obtained from the left and right singular vectors of the associated matrix.The update uses an unfolding of the tensor and the contraction of the other factors.Source: C. Low-Rank MERA Approximation · p. 4
• Each isometry is updated by minimizing reconstruction error subject to an orthonormality constraint on its unfolding, using a solution analogous to the disentangler update.The two isometry subproblems use different mode unfoldings of the tensor.Source: C. Low-Rank MERA Approximation · p. 4
• With the layer factors fixed, the top-core subproblem minimizes reconstruction error and is updated by contracting the tensor with the layer.The stated update relies on the orthogonality of the disentangler and orthonormality of the isometries.Source: C. Low-Rank MERA Approximation · p. 4

A. Low-Rank MERA Approximation for Compression, Denoising and Feature Extraction

• A 3D OCT volume is transformed into a 5D tensor and approximated by a top core, a disentangler, and two isometries.The resulting factors represent the compressed volume.Source: A. Low-Rank MERA Approximation for… · p. 4
• The compression ratio is defined by the data tensor size relative to the MERA decomposition size, and determines the MERA rank.The rank is obtained by solving a quadratic equation and rounding the solution to an integer.Source: A. Low-Rank MERA Approximation for… · pp. 4–5
• The estimated MERA factors form the compressed volume; their reconstruction produces a low-rank approximation of the original OCT volume.The reconstruction computes a layer tensor and the top core before forming the volume estimate.Source: A. Low-Rank MERA Approximation for… · p. 5
• The top core encodes global features carrying inter- and intra-B-scan information, while the compression ratio controls the number of features.The denoised volume estimate can be used for retinal-layer segmentation and visual diagnosis.Source: A. Low-Rank MERA Approximation for… · p. 5

B. Feature Extraction

• The method assumes diagnosis-specific information in a 3D OCT volume is contained in the top core G.The feature vector is formed by vectorizing the entrywise absolute values of G.Source: B. Feature Extraction · p. 5
• MERA feature vectors are grouped according to class, with N_c denoting the number of OCT volumes in class c.The grouping follows the feature-vector definition.Source: B. Feature Extraction · p. 5
• The section proposes GM-NSC for cases with limited labeled training samples, where deep learning can have reduced generalization performance.The passage identifies rare retinal diseases as an example of this setting,.Source: B. Feature Extraction · p. 5

C. Grassmann-Manifold Nearest Subspace Classifier

• The method models feature data through a union of low-dimensional subspaces, motivated by data having fewer degrees of freedom than their ambient domain.Each class subspace is treated as a point on the Grassmann manifold, the set of d-dimensional linear subspaces of a D-dimensional Euclidean space.Source: C. Grassmann-Manifold Nearest Subspace… · p. 5
• The classifier estimates class representatives from training feature vectors and uses a minimum point-to-subspace distance criterion to label an unlabeled vector.The subspace dimension d is identified as unknown in the classification description.Source: C. Grassmann-Manifold Nearest Subspace… · p. 5
• The outlined pipeline maps each 3D OCT volume to a 5D tensor, performs low-rank MERA approximation, extracts and groups features, estimates Grassmann points, and assigns a diagnosis.The algorithm describes the output as a diagnosis for the specific test volume.Source: C. Grassmann-Manifold Nearest Subspace… · p. 6

D. Generalization in a small-sample scenario

• GM-NSC has the subspace dimension d as its only hyperparameter.The section links its small-sample rationale to the accuracy of estimated subspace bases.Source: D. Generalization in a small-sample… · p. 6
• Basis-approximation error can be reduced when subspace dimension d is small relative to D; in age-related macular degeneration and glaucoma grading, d is never greater than 12 and is quite often 1.Source: D. Generalization in a small-sample… · p. 6
• The authors state that, under this small-dimension condition, label assignment should remain reasonably accurate and GM-NSC should retain good generalization.Their argument refers to subspace basis estimation under noise and an eigenvalue gap.Source: D. Generalization in a small-sample… · p. 6

• The section reports MERA results for OCT compression, denoising, feature extraction, and retinal-disease diagnosis.Source: §V · p. 6

A. Data and Code

• The data comprise Topcon and in-house BV1000 OCT volumes, and the authors provide MATLAB and PyTorch code with demos.Topcon volumes contain 64 B-scans of 480 × 512 pixels; BV1000 volumes contain 100 B-scans of 1000 × 512 pixels.Source: A. Data and Code · p. 6

B. Segmentation of Retinal Layers and Improvement of Visual Quality

• Ophthalmologists graded three B-scans from three OCT volumes per scanner for the expert image-quality assessment.EIQM is reported as mean ± standard deviation; the supplied passage begins, but does not complete, its grade-scale description.Source: B. Segmentation of Retinal Layers and… · p. 7
• TOPCON segmentation errors were 1.88% at CR=4 and 1.87% at CR=10; BVS errors stayed below 4% at CR=50 and near 3% for CR≤30.For BVS, these results compare fairly with no-compression median and bilateral filtering; at CR=40, median filtering slightly outperformed MERA.Source: B. Segmentation of Retinal Layers and… · pp. 6–7, Figure 2 · p. 7
• For TOPCON at CR≤30, MERA's EIQM compared fairly with the original and low-rank TT, median-filtered, and bilateral-filtered images, but trailed the original and low-rank MLR.For BVS at CR≤40, MERA and low-rank MLR highly improved visual quality; median and bilateral filtering were no-compression methods.Source: Figure 3 · p. 7, B. Segmentation of Retinal Layers and… · p. 7

C. Feature Extraction and Diagnosis of Retinal Diseases

• For AMD, peak accuracy occurred at subspace dimensions 12, 11, 10, 6, 6, and 6 for compression ratios 1.6, 2, 4, 10, 20, and 40, respectively.The authors interpret these peaks as supporting low-dimensional representations of the high-dimensional MERA features.Source: C. Feature Extraction and Diagnosis of… · p. 7, Figure 4 · p. 8
• On AMD features, GM-NSC outperformed optimized SVM and NN classifiers, trained substantially faster, and matched ResNet18 accuracy especially at smaller compression ratios.GM-NSC training was ten to hundred times faster than SVM and more than three orders of magnitude faster than NN.Source: C. Feature Extraction and Diagnosis of… · p. 8, Table I · p. 8
• For glaucoma grading, GM-NSC mostly outperformed SVM with roughly two orders of magnitude lower computational complexity, while DCNN led except in early-versus-advanced grading.With only 54/50 training samples in that scenario, DCNN performed worse than GM-NSC.Source: C. Feature Extraction and Diagnosis of… · pp. 8–9

VI. CONCLUSION

• The conclusion reports improved visual quality and reduced retinal-layer segmentation error at compression ratios up to 40, alongside low-dimensional MERA features for diagnosis.The image-quality result was especially evident on images from the low-quality scanner.Source: §VI · p. 9