Focus-to-Perceive Representation Learning: A Cognition-Inspired Hierarchical Framework for Endoscopic Video Analysis
A new AI mimics how doctors examine tissue, focusing first on static lesions before tracking their evolution.
A research team has introduced Focus-to-Perceive Representation Learning (FPRL), a novel AI framework designed to overcome a critical bottleneck in medical AI: the lack of high-quality annotated data for endoscopic video analysis. Published on arXiv and accepted to CVPR 2026, the work addresses a flaw in standard self-supervised video models, which are built for natural scenes and exhibit a 'motion bias.' This bias causes them to prioritize moving elements, overlooking the static, lesion-centric visual semantics that are paramount for clinical decision-making in gastrointestinal screening.
FPRL innovates by mimicking a clinician's hierarchical examination process. The framework first 'Focuses,' using a Teacher-Prior Adaptive Masking (TPAM) technique to concentrate the model on intra-frame, lesion-relevant regions to learn static semantics. It then 'Perceives,' employing modules like Cross-View Masked Feature Completion (CVMFC) to model how these regions evolve contextually across frames. This explicit separation and collaborative learning of static and contextual semantics allows the AI to build more medically relevant video representations without dense manual labels.
The results are compelling. In extensive validation across 11 different endoscopic video datasets, FPRL demonstrated superior performance on diverse downstream tasks compared to existing pre-training methods. By effectively filtering out redundant temporal noise and capturing structured semantic evolution, the framework provides a more robust foundation for building diagnostic tools. The code has been made publicly available, paving the way for more accurate, data-efficient AI assistants in endoscopy.
- Mimics clinical reasoning with a two-stage 'Focus then Perceive' hierarchy, first analyzing static lesions, then their temporal evolution.
- Introduces novel techniques like Teacher-Prior Adaptive Masking (TPAM) to reduce motion bias and prioritize clinically salient image regions.
- Outperforms existing video AI methods across 11 diverse medical datasets, proving effectiveness for tasks like polyp detection and classification.
Why It Matters
This enables more accurate, label-efficient AI diagnostic tools for early cancer detection in gastrointestinal endoscopy, a critical screening procedure.