Reproduce “EndoNet: A Deep Architecture for Recognition Tasks on Laparoscopic Videos” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Andru P. Twinanda, Sherif Shehata, Didier Mutter, Jacques Marescaux, Michel de Mathelin, Nicolas Padoy
EndoNet, developed by researchers at the University of Strasbourg and its affiliated medical institutes, introduces a deep learning architecture that simultaneously performs surgical phase recognition and tool presence detection solely from laparoscopic video frames. It achieved 92.0% accuracy for offline phase recognition and 81.0% mean Average Precision for tool detection on the custom Cholec80 dataset, demonstrating the effectiveness of multi-task learning for surgical workflow analysis.
Key insights
- Multi-task learning, specifically combining tool presence detection with surgical phase recognition, significantly enhances the discriminative power of learned visual features for both tasks.
- Deep learning models, especially CNNs fine-tuned on domain-specific data, can effectively learn robust visual features directly from raw laparoscopic video, overcoming the limitations of handcrafted features and reducing reliance on manual annotations or external sensors.
- Temporal modeling (e.g., using Hidden Markov Models) is crucial for achieving high accuracy and robust performance in surgical phase recognition, even when strong frame-level features are available.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.