Exploring the challenges of artifact detection in endoscopic videos through the perspective of long-tailed learning.

Through analysis of artifact distribution characteristics in endoscopic videos from the public EAD2020 dataset, we develop a long-tailed endoscopic artifact detection algorithm via a decoupled framework. The proposed multi-expert system strategically partitions head and tail classes into two distinct subsets, with domain-specific experts processing each group separately. We innovatively design a hierarchical expert aggregation mechanism and further integrate visual prompts to enhance cross-group discrimination capabilities, ultimately achieving a significant improvement in artifact detection performance as measured by the comprehensive mean Average Precision (mAP) metric.

Benefiting from a more powerful backbone network and advanced classifier design, the proposed method achieves significant improvements in endoscopic video analysis: mAP for sequential video frame detection increases from 0.440 to 0.491, while mAP for individual video frame detection rises from 0.339 to 0.384. Notably, our approach comprehensively outperforms existing ensemble methods across all artifact categories.

The proposed multi-expert framework for long-tailed endoscopic video artifact detection achieves balanced performance across artifact categories through grouping and multi-expert methods. This architecture significantly enhances detection performance for tail classes while maintaining head class accuracy, effectively mitigating the inherent class imbalance problem. This study provides a novel direction for medical image detection tasks, particularly in data-imbalanced scenarios.