Reproduce “SurgPETL: Parameter-Efficient Image-to-Surgical-Video Transfer Learning for Surgical Phase Recognition” on OpenResearch

About this paper

Shu Yang, Zhiyuan Cai, Luyang Luo, Ning Ma, Shuchang Xu, Hao Chen

Researchers from HKUST developed SurgPETL, a parameter-efficient image-to-surgical-video transfer learning framework incorporating a Spatial-Temporal Adaptation (STA) module, to specialize image-level pre-trained models for fine-grained surgical phase recognition. This approach achieved state-of-the-art performance on multiple surgical video datasets, such as a 6.2% Jaccard gain on AutoLaparo over previous methods, while fine-tuning substantially fewer parameters.

Key insights

• Pre-training on surgical domain datasets significantly enhances performance compared to natural image datasets, indicating the importance of domain-specific feature learning for transferability.
• Parameter-efficient transfer learning methods can outperform full fine-tuning on surgical video tasks while fine-tuning substantially fewer parameters, suggesting improved efficiency without sacrificing accuracy.
• Explicitly designed temporal adaptation, like the proposed Feature Re-embedding, is critical for capturing fine-grained inter-frame dynamics necessary for accurate surgical phase recognition, where general PEIVTL methods fall short.

Summary by alphaXiv.