Retinal fundus images can show signs of multiple diseases at once, making classification a multi-label problem. Public datasets for this task are heterogeneous and severely long-tailed, according to a new paper on arXiv.
The paper observes that clinical triage often implies a staged pipeline, starting with a screening gate and then a disease-group router. The authors argue for an alternative: a label-correlation ensemble that captures relationships between diseases instead of relying on sequential stages.
The proposed approach is positioned as "beyond multi-stage pipelines," suggesting that modeling label correlations directly may better handle the complexity and imbalance of real-world fundus data. The provided abstract does not include experimental results or detailed architecture, so the claims are limited to the proposed design.