[Paper Review] Gradient based Severity Labeling for Biomarker Classification in OCT

Kiran Kokilepersaud, Mohit Prabhushankar+3 more|arXiv (Cornell University)|Feb 23, 2026

Retinal Imaging and Analysis0 citations

TL;DR

The paper introduces a severity-based pseudo-labeling strategy for unlabeled OCT scans using gradient responses from a healthy auto-encoder, and trains a supervised contrastive encoder to improve biomarker classification in OCT.

ABSTRACT

In this paper, we propose a novel selection strategy for contrastive learning for medical images. On natural images, contrastive learning uses augmentations to select positive and negative pairs for the contrastive loss. However, in the medical domain, arbitrary augmentations have the potential to distort small localized regions that contain the biomarkers we are interested in detecting. A more intuitive approach is to select samples with similar disease severity characteristics, since these samples are more likely to have similar structures related to the progression of a disease. To enable this, we introduce a method that generates disease severity labels for unlabeled OCT scans on the basis of gradient responses from an anomaly detection algorithm. These labels are used to train a supervised contrastive learning setup to improve biomarker classification accuracy by as much as 6% above self-supervised baselines for key indicators of Diabetic Retinopathy.

Motivation & Objective

• Motivate accurate biomarker detection in OCT with limited labeled data.
• Develop a severity-based pseudo-labeling mechanism using gradient responses.
• Leverage supervised contrastive learning to improve biomarker classification performance.
• Evaluate how severity-bin choices affect biomarker detection across multiple OCT biomarkers.

Proposed method

• Train an auto-encoder on healthy OCT to align gradients across healthy samples using GradCON.
• Compute a severity score SS = -L_recon + α L_grad for unlabeled OCT images, with α = 0.03.
• Discretize severity scores into N bins to assign severity labels SL to unlabeled data.
• Train a ResNet-18 encoder with supervised contrastive loss using severity labels.
• Freeze encoder weights and fine-tune a linear classifier on labeled biomarker data.

Fig. 1 : OCT Scans of biomarkers from the Prime + TREX DME datasets. The biomarkers are A) Intraretinal Hyperreflective Foci (IRHRF), B) Intraretinal Fluid (IRF) and Diabetic Macular Edema (DME) C) Partially Attached Vitreous Face (PAVF), and D) Fully Attached Vitreous Face (FAVF). The white arrows

Experimental results

Research questions

• RQ1 Can gradient-based severity labels from healthy distributions improve representation learning for OCT biomarkers?
• RQ2 How does the number of severity bins N affect multi-label AUC and per-biomarker performance?
• RQ3 Do severity-based supervised contrastive representations outperform self-supervised baselines on OCT biomarker tasks?
• RQ4 What is the impact of using different anomaly detectors to generate severity labels?

Key findings

• Severity-label based supervised contrastive learning yields improved mean multi-label AUC and biomarker accuracy over self-supervised baselines.
• Varying the severity bin count N changes performance; moderate N (e.g., 5000–10000) often yields better multi-label AUC than extreme values.
• Different anomaly detectors for severity labeling show SL (severity-label) method generally outperforms others in multi-label results.
• Best per-biomarker performance depends on the biomarker, with some benefiting from higher bin counts (e.g., DME, IRF) and others from moderate counts.

Fig. 2 : From a healthy manifold learned from a trained auto-encoder, we can compute distance to the manifold of more severely diseased cases via a Severity Score (SS). Severity Score is calculated via some model response and increases as a sample is more anomalous compared to the learned healthy ma