Available View Set-guided multi-view retinal disease classification from OCT and OCTA

Optical coherence tomography (OCT) and OCT angiography (OCTA) are pivotal technologies in retinal imaging. Clinically, a comprehensive assessment of retinal diseases relies on integrating structural and vascular features extracted from layer-specific projections. However, this integration is inevitably constrained by the Available View Set-the slides that are obtainable under the current scanning mode and remain interpretable after automated layering. To this end, we propose the Adaptive Multi-View Transformer with Uncertainty (AMTU), designed to exploit complementary information while reducing the discriminative dependency on the integrity of the Available View Set. Specifically, this model maximizes complementary pathological information from different views through cross-layer joint modeling, while employing an availability-normalized fusion and uncertainty-driven weighting mechanism to prioritize high-quality views. Furthermore, an uncertainty-regulated class-center alignment mitigates feature shifts via high-quality view constraints, maintaining stable classification representations under varying availability conditions. Experiments on the OCTA-500 dataset for the classification of four diagnostic categories showed that AMTU achieves a balanced accuracy of 75.9% under full available inputs and demonstrates exceptional robustness when the available view set is constrained. Overall, our model provides a unified solution for diverse clinical scenarios, with its precise classification offering a reliable basis for clinical diagnosis.