[Paper Review] Classification of dry age-related macular degeneration and diabetic macular edema from optical coherence tomography images using dictionary learning

Elahe Mousavi, Rahele Kafieh+1 more|arXiv (Cornell University)|Mar 16, 2019

Retinal Imaging and Analysis4 citations

TL;DR

This paper proposes a dictionary learning-based classification framework for distinguishing dry age-related macular degeneration (AMD), diabetic macular edema (DME), and normal cases from optical coherence tomography (OCT) images without relying on retinal layer segmentation. By extracting histogram of oriented gradients (HOG) features and applying Fused Dictionary Learning (FDDL), the method achieves 95.13% accuracy for AMD, 100% for DME, and 100% for normal cases using only 4% of B-scans per volume, demonstrating high sensitivity to early-stage disease signs.

ABSTRACT

Age-related Macular Degeneration (AMD) and Diabetic Macular Edema (DME) are the major causes of vision loss in developed countries. Alteration of retinal layer structure and appearance of exudate are the most significant signs of these diseases. With the aim of automatic classification of DME, AMD and normal subjects from Optical Coherence Tomography (OCT) images, we proposed a classification algorithm. The two important issues intended in this approach are, not utilizing retinal layer segmentation which by itself is a challenging task and attempting to identify diseases in their early stages, where the signs of diseases appear in a small fraction of B-Scans. We used a histogram of oriented gradients (HOG) feature descriptor to well characterize the distribution of local intensity gradients and edge directions. In order to capture the structure of extracted features, we employed different dictionary learning-based classifiers. Our dataset consists of 45 subjects: 15 patients with AMD, 15 patients with DME and 15 normal subjects. The proposed classifier leads to an accuracy of 95.13%, 100.00%, and 100.00% for DME, AMD, and normal OCT images, respectively, only by considering the 4% of all B-Scans of a volume which outperforms the state of the art methods.

Motivation & Objective

• To develop an automated, segmentation-free method for classifying dry AMD, DME, and normal OCT images.
• To detect early-stage retinal pathologies where disease signs appear in only a small fraction of B-scans.
• To improve classification accuracy over state-of-the-art methods by leveraging discriminative feature learning without retinal layer segmentation.
• To demonstrate the feasibility of early diagnosis using minimal representative B-scans.

Proposed method

• Extract histogram of oriented gradients (HOG) features from OCT B-scans to capture local intensity gradients and edge structures.
• Apply three dictionary learning methods—Fused Dictionary Learning (FDDL), Collaborative Representation-based Dictionary Learning (COPAR), and Low-Rank Sparse Dictionary Learning (LRSDL)—to learn discriminative features.
• Use a leave-three-out cross-validation strategy on a dataset of 45 subjects (15 per class) to train and test the classifiers.
• Classify each 3D OCT volume based on the majority vote of B-scan predictions, with final labels determined by the most frequent class among B-scans.
• Optimize the framework to detect disease from as little as 4% of B-scans, emphasizing early diagnosis.
• Evaluate performance using accuracy, and compare with existing methods using identical datasets and cross-validation protocols.

Experimental results

Research questions

• RQ1 Can a dictionary learning-based classifier achieve high accuracy in distinguishing AMD, DME, and normal OCT images without requiring retinal layer segmentation?
• RQ2 Can the proposed method detect early-stage AMD and DME when pathological signs are present in only a small fraction of B-scans?
• RQ3 How does the performance of FDDL, COPAR, and LRSDL compare when applied to HOG features for OCT image classification?
• RQ4 Can the method achieve superior accuracy compared to state-of-the-art approaches using minimal B-scan representation?
• RQ5 What is the minimum percentage of B-scans required for reliable classification, indicating early disease detection capability?

Key findings

• The proposed method achieved 95.13% accuracy for AMD, 100.00% for DME, and 100.00% for normal cases using only 4% of B-scans per volume.
• Fused Dictionary Learning (FDDL) with HOG features outperformed COPAR and LRSDL, achieving the highest overall accuracy of 98.37% on the full dataset.
• The method successfully detected all DME and AMD cases, outperforming Srinivasan et al. and Wang Yu et al. in normal case classification.
• The FDDL-based classifier required only 22 minutes for leave-three-out cross-validation, demonstrating efficient training compared to COPAR (41 min) and LRSDL (32 min).
• The approach achieved superior performance compared to prior methods, especially in detecting normal cases and early-stage disease, due to its robust feature representation and minimal reliance on full-volume analysis.
• The results confirm that dictionary learning with HOG features enables reliable, early-stage detection of retinal pathologies even when pathological signs are sparse in the B-scan volume.