🔎 Gemini leads GPT in OCT feature F1, 0.558 vs 0.472
🔎 Gemini leads GPT in OCT feature F1, 0.558 vs 0.472
In a cross-sectional study of 330 publicly available retinal OCT B-scans across 12 diagnostic categories, Gemini 3 Pro outperformed GPT 5.2 Pro on both OCT feature detection (macro-F1 0.558 vs. 0.472) and diagnostic accuracy (58.8% vs. 32.1%). The study suggests that image interpretation alone is no longer enough for safe patient-facing OCT triage: both models frequently overcalled pathology, missed subtle features, and showed recurrent diagnostic biases.
Why It Matters To Your Practice
Patients can now access multimodal AI tools that interpret retinal images, increasing the odds that clinicians will need to address AI-generated OCT readouts in routine care.
Both models were prone to false positives on normal scans, raising concern for unnecessary anxiety, inappropriate referrals, and extra follow-up testing.
Performance was stronger for visually obvious abnormalities than for subtle structural findings, limiting reliability in cases where nuance matters most.
Clinical Implications
Do not treat current general-purpose LMM OCT interpretations as a stand-alone diagnostic aid or screening tool without specialist oversight.
If AI outputs enter workflow, consider using them only as a secondary support layer with explicit guardrails, not as a substitute for clinician interpretation.
Be especially cautious with “normal vs. abnormal” calls: GPT 5.2 Pro did not correctly identify any normal images in this dataset.
Insights
The study separately tested nine predefined OCT features and free-text diagnoses using zero-shot inference, offering a practical look at out-of-the-box performance.
Gemini 3 Pro led GPT 5.2 Pro not just on feature extraction but also on diagnosis, suggesting that better visual feature capture may translate into better overall interpretation.
Still, absolute performance remained modest, and hallucinated diagnoses plus model-specific error patterns point to systematic—not random—failure modes.
The Bottom Line
Gemini currently performs better than GPT on retinal OCT interpretation in this study, but neither model appears reliable enough for unsupervised clinical or patient-facing use.
For clinicians, the near-term role of these tools is likely limited to tightly supervised assistance, with local validation and clear communication about error risk.