SMASH 2026 · Paper #21 · Result card
Two for One: Privileged Bilateral Learning for Monocular Retinal Disease Classification
The student beat the single-eye baseline in 8 of 10 simulated cohorts (mean +0.024 AUROC). On the one split shown in detail it scored slightly lower (0.711 vs 0.702).
task preset amdeye at inference leftprevalence 0.06inter-eye correlation 0.6α 0.03eye noise L / R 2.0 / 2.0patients 3500cohorts 10data toy simulated paired-eye features (not ODIR images)
AUROC, left eye only
0.711 → 0.702
baseline → privileged student; Δ 95% CI [-0.033, +0.012] (1,000 bootstraps)
Robust across cohorts?
8 of 10
independent cohorts/splits where AUROC improved; mean Δ +0.024 ± 0.021
Trade-off at the macro-F1 threshold
FPR 0.027 → 0.047
precision 0.357 → 0.244; recall 0.238 → 0.238
Student ↔ teacher similarity
ΔCKA +0.118
linear probe AUROC 0.718 → 0.706; teacher (both eyes) 0.729
From the abstract
We thus propose a privileged learning framework in which a teacher model has access to bilateral CFP images during training and learns representations informed by information from both eyes, while the student is trained to transfer this knowledge from the bilateral teacher using only a single-eye image as input. At inference, the student operates exclusively on this single-eye input and does not require the contralateral image.
Evidence
Paper vs. your run
| Metric | Paper | Your run | |
|---|---|---|---|
| AUROC, baseline → privileged | AMD left 0.931 → 0.966; AMD right 0.801 → 0.810; DR left 0.789 → 0.836; DR right 0.825 → 0.837 | 0.711 → 0.702 (left, amd preset) | toy: magnitudes are not comparable |
| Balanced accuracy | AMD left 0.834 → 0.879; DR left 0.576 → 0.749 | 0.605 → 0.595 | |
| FPR / precision trade-off | AMD left FPR 0.011 → 0.027, precision 0.760 → 0.595; DR left FPR 0.766 → 0.089, precision 0.078 → 0.760 | FPR 0.027 → 0.047, precision 0.357 → 0.244 | both directions happen; depends on the validation threshold |
| ΔCKA (student vs baseline, to teacher) | +0.015 (DR); AMD inconclusive | +0.118 | linear CKA, validation set |
| Linear probe AUROC | AMD 0.815 vs 0.836 (+2.0 points); DR +2.9 points | 0.718 vs 0.706 (-1.2 points) |
Try next
- Eyes nearly identical — The sweep shows the gain falling to +0.014 at inter-eye correlation 0.95; this checks whether a second eye that agrees with the first adds less.
{"inter_eye_corr": 0.95} - Noisy deployment eye, clean privileged eye — The left eye is noisier at inference while the teacher gets a cleaner right eye; tests whether the privileged eye helps more when the deployed image is poor.
{"noise_left": 3.0, "noise_right": 1.0} - Common disease (DR preset) — Same settings with 30% prevalence; compares the detailed-split gain and the FPR and precision trade-off with AMD.
{"task": "dr"}
Read this before sharing
- Simulated eye features replace fundus photographs.
- A small MLP replaces RETFound-Green.