TL;DR
RetiDual-Select is a reliability-aware framework that intelligently routes retinal OCT images between a domain specialist (RETFound) and a medical generalist (MedSigLIP). By introducing a validation-frozen "harm guard," the system ensures that adaptive model switching only occurs when it's likely to repair errors, significantly improving robustness to image-quality drops like severe contrast reduction.
What Problem Does This Paper Address?
Deep learning has revolutionized retinal Optical Coherence Tomography (OCT) analysis, but a critical challenge remains: reliability under distribution shift. When image quality degrades (due to noise, blur, or low contrast), even the best models can fail.
Researchers often use ensembles—combining multiple models—to improve performance. A common approach is "Static Product-of-Experts" (PoE), which averages model outputs. While robust, PoE can't exploit the fact that one model might be much better than another for a specific, degraded image. "Adaptive routing" attempts to pick the best model for each image, but it's a double-edged sword: a wrong routing decision can overturn a correct ensemble prediction, introducing new errors. This paper tackles the problem of making adaptive fusion safe and auditable.
Methods
The authors propose RetiDual-Select, a hierarchical decision architecture (illustrated in Figure 1) that separates three distinct tasks:
- Candidate Generation: Two frozen encoders—RETFound (a retinal specialist) and MedSigLIP (a medical generalist)—produce calibrated class probabilities. These are combined into two candidates: a conservative Static PoE and an Adaptive Gate output.
- Routing Safety (The Harm Guard): This is the core innovation. A "harm guard" model estimates the risk that the adaptive gate might perform worse than the static PoE. If the risk is too high, the system falls back to the safer Static PoE.
- Predictive Review: Finally, a separate risk score determines if the selected prediction is certain enough to be released automatically or if it warrants human review.
Datasets
The study evaluates the framework across three major OCT benchmarks:
- Kermany-4: A standard 4-class dataset (~50k images).
- OCT-C8: An 8-class dataset (~24k images).
- OCTDL: A 7-class dataset (~2k images) used for independent-task validation.
- Synthetic Quality Suite: To test robustness, the authors applied 15 types of controlled degradations (Gaussian blur, noise, contrast/brightness shifts, and JPEG compression) at three severity levels.
Key Insights and Ideas
- Static PoE is a formidable baseline: On clean data, the simple static ensemble reached the highest accuracy, outperforming both individual models and the adaptive gate. This establishes PoE as the "conservative reference" that any adaptive system must respect.
- The "Opportunity-Alignment" Framework: For routing to be useful, two things must happen: there must be a reliability opportunity (one model is significantly more robust to a specific degradation than the other) and the gate must have alignment (it must correctly identify and move toward the more reliable model).
- Specialists are sensitive to contrast: A major "aha moment" was discovering that the specialist (RETFound) is much more sensitive to contrast reduction than the generalist (MedSigLIP). This creates a massive opportunity for adaptive routing in low-contrast scenarios (see Figure 6).
- Safety through Fallback: By treating routing as a safety problem rather than just an accuracy problem, the authors show that a "harm guard" can prune away risky adaptive decisions, retaining the benefits of routing while mitigating the downsides.
Key Results
- Clean Performance: Static PoE achieved 98.76% accuracy on Kermany-4 and 96.42% on OCT-C8.
- Adaptive Gains: Under severe contrast reduction, adaptive routing provided massive improvements over the static reference: +4.65 percentage points (pp) on Kermany-4 and +19.88 pp on OCT-C8.
- Harm Guard Efficacy: The validation-frozen harm guard successfully reduced the "worst-case" loss caused by bad routing. On OCT-C8, it cut the maximum routing error from 1.39 pp down to 0.36 pp.
- Selective Accuracy: At a 2% error target, the system's review policy accepted over 92% of clean images while maintaining a selective accuracy of 98.65% to 99.89%.
Limitations and Future Work
- Synthetic vs. Natural Shift: The study relies on synthetic perturbations. While these are useful for "mechanism audits," they don't perfectly replicate natural shifts caused by different scanner brands or clinical settings.
- Dataset Integrity: The authors discovered a significant overlap (31.74% byte-identical images) between the OCT-C8 and Kermany-4 datasets, highlighting the need for more rigorous data auditing in the field.
- Future Directions: The next step is prospective validation using real-world, low-quality acquisitions from multiple clinical centers to see if the "specialist-generalist" advantage holds up in the wild.
Summary updated