Technical Summary: Optic Disc Colour Ratio (ODCR) for Subgroup Fairness in Diabetic Retinopathy Screening

Problem Statement

Automated Diabetic Retinopathy (DR) screening systems achieve specialist-level accuracy on curated benchmarks but exhibit poorly understood performance disparities across subgroups. While prior work often attributes these disparities to data imbalance, this paper argues for a mechanistic explanation: choroidal pigmentation (melanin concentration) reduces the optical contrast between lesions (microaneurysms, hemorrhages) and the retinal background.

This reduction in Signal-to-Noise Ratio (SNR) physically degrades the input signal available to convolutional lesion detectors. The authors hypothesize that this leads to a specific failure mode: over-prediction. Darker-pigmented fundi, having lower lesion-background contrast, are systematically flagged as "referable" at higher rates regardless of their true label, imposing a disproportionate false-positive burden on underserved populations. This contrasts with dermatological AI, where similar contrast issues often manifest as specificity failures (missed detections).

Methodology

1. The ODCR Proxy

To audit fairness without relying on ground-truth ethnicity labels, the authors introduce the Optic Disc Colour Ratio (ODCR).

• Definition: A metric computed from the peripapillary annulus (the region surrounding the optic disc) in CIE $L^*a^*b^*$ color space.
• Formula: $odcr = \arctan(\frac{\bar{L}^* - 50}{\bar{b}^*}) \times \frac{180}{\pi}$.
• Function: Higher values indicate lighter pigmentation; lower values indicate darker pigmentation. It serves as a proxy analogous to the Individual Typology Angle (ITA) used in dermatology.
• Stratification: Patients are grouped into "Dark," "Medium," and "Light" based on ODCR tertiles from the training set.

2. SNR Analysis

The authors quantify the vessel-background SNR to validate the physical mechanism.

• Metric: $SNR = \frac{|\mu_{lesion} - \mu_{background}|}{\sigma_{background}}$.
• Method: Candidate microaneurysm pixels are identified via adaptive green-channel thresholding (peak hemoglobin absorption), while background is sampled from mid-luminance peripapillary tissue.

3. Experimental Setup

• Datasets:
  • EyePACS: 35,126 images (US-based). Patient-level splits were used (Train: 24,588; Val: 5,268; Test: 5,270) to prevent data leakage.
  • APTOS 2019: 3,662 images (Indian population) used for zero-shot transfer validation.
• Model Architecture: EfficientNet-B2 backbone with a conditioning branch (Linear3→16→32) that concatenates tone embeddings.
• Ablation Study: Five variants were tested on the locked EyePACS test set:
  1. Baseline: Standard training ($pos\_weight=2.0$).
  2. Aug-only: Dark fundus augmentation + 3x oversampling.
  3. Tone-only: ODCR conditioning.
  4. Full: Combination of all interventions.
  5. Balanced-baseline: Equalized sampling and $pos\_weight=1.0$ (control).
• Evaluation: Metrics included AUC, Sensitivity, and Specificity, stratified by ODCR groups. Statistical significance was assessed via permutation tests and bootstrap confidence intervals.

Key Results

1. SNR Reduction in Dark Fundi

The study confirms a significant group-level reduction in lesion-background contrast for darker fundi.

• Magnitude: The light-to-dark SNR ratio is 1.14×.
• Significance: This reduction is statistically significant ($p = 3 \times 10^{-10}$ via Kruskal–Wallis; $p = 0.0005$ via Mann–Whitney).
• Comparison: This effect is attenuated compared to dermatology (where the ratio was 5.2×), likely because the choroid lies behind the retinal pigment epithelium and active fundus illumination partially compensates for absorption.

2. The "Score-Elevation Signature"

Across all five model variants, dark-pigmented fundi received systematically higher predicted probabilities than light-pigmented fundi, regardless of the true label.

• Evidence: In the Baseline variant, dark true-negatives had a mean score of 0.128 vs. 0.105 for light true-negatives ($\Delta = +0.023$).
• Implication: This indicates global model uncertainty under low contrast, leading to a shift toward the positive class.

3. Sensitivity Asymmetry (Over-Prediction)

Under sensitivity-calibrated training (specifically the Full variant), the score elevation manifests as a significant sensitivity gap rather than a specificity gap.

• Full Variant Gap: Dark sensitivity (0.642) vs. Light sensitivity (0.535), $\Delta = +0.107$ ($p = 0.015$).
• Specificity Gap: The specificity gap was small and non-significant ($\Delta \approx -0.009$).
• Clinical Impact: This results in approximately 4 excess referrals per 1,000 dark-pigmented patients compared to light-pigmented patients.
• AUC Insufficiency: The AUC gaps between groups were negligible ($\le 0.008$) and non-significant, confirming that AUC alone is insufficient for fairness auditing in this context.

4. External Validation

Zero-shot transfer to the APTOS 2019 dataset (Indian population) replicated the signature. The score elevation was even more pronounced ($\Delta = +0.101$ for true-negatives), consistent with the darker population distribution.

5. Causal Structure

• Threshold Calibration: Adjusting decision thresholds ($t = 0.30–0.65$) did not close the gap, indicating the disparity is embedded in the predicted probability distributions, not just the decision boundary.
• Balancing: Class balancing reduced the gap relative to the "Full" variant but did not eliminate it below the "Baseline" level. This suggests a mixed causal structure: the disparity is driven substantially by the sensitivity-calibrated decision rule (oversampling and high $pos\_weight$) interacting with the physical SNR deficit.

Significance and Claims

The paper claims to provide a mechanistic understanding of fairness disparities in DR screening, moving beyond data imbalance to physical optical properties.

1. ODCR as a Proxy: It establishes ODCR as a computationally accessible, label-free proxy for fundus pigmentation, enabling subgroup analysis without ethnicity labels.
2. SNR as a Predictor of Disparity Type: The paper proposes that the magnitude of the SNR ratio may predict the axis of the disparity.
   • High SNR reduction (Dermatology, 5.2×) $\rightarrow$ Specificity failure (under-detection).
   • Attenuated SNR reduction (Fundus, 1.14×) $\rightarrow$ Sensitivity failure (over-prediction).
3. Limitations of AUC: It demonstrates that standard metrics like AUC can mask significant subgroup disparities in DR screening, necessitating stratified sensitivity/specificity reporting.
4. Clinical Recommendation: The authors argue that fairness auditing must focus on over-prediction (excess false positives) in dark-pigmented populations and that post-hoc threshold calibration is insufficient to resolve these disparities.

The study concludes that choroidal pigmentation is a measurable, statistically significant group-level confounder that drives over-prediction in DR screening, a phenomenon that is invisible to AUC but detectable through ODCR-stratified analysis.