Occluded Instrument Feature Extraction via Multi-Scale Attentional Networks with Receptive Field Enhancement

Endoscopic retrograde cholangiopancreatography (ERCP) is a crucial technique for diagnosing and treating biliary and pancreatic diseases. With the rapid advancement of embodied intelligence and robotic technology, autonomous robotic intervention imposes higher demands on the precise extraction of duodenal papilla key points. However, existing methods still face significant challenges due to instrument occlusion and extreme viewing angles. To address these issues, we propose a series of optimization strategies for the YOLOv11 algorithm, including integrating our MRF-Fusion module at the Backbone’s backend, incorporating our MOaE-Fusion layer in the Neck, and refining the network connection strategy to enhance feature extraction and multi-scale occlusion perception. Additionally, we construct ODT-60, a dedicated dataset for evaluating occlusion and distortion robustness. Experimental results demonstrate that our improved model achieves state-of-the-art performance on ODT-60, with key point recall improving by up to 33.3% (reaching 95%) and mAP50 increasing by up to 24.3% (reaching 94.8%). These findings provide a more precise and stable key point extraction solution for autonomous endoscopic robots, with real-time performance, ensuring accurate processing. This lays a robust foundation for the further development of intelligent biliary and pancreatic endoscopic navigation systems.