SurgPETL: Parameter-Efficient Image-to-Surgical-Video Transfer Learning for Surgical Phase Recognition

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Machine Learning

Background:

• Image pre-training shows promise for downstream tasks, but video fine-tuning faces bottlenecks.
• Surgical video analysis is challenged by limited data and the need for spatiotemporal modeling.
• Parameter-Efficient Image-to-Video Transfer Learning (PEIVTL) offers efficiency but its surgical application is unexplored.

Purpose of the Study:

• To develop an efficient method for adapting image-level pre-trained models for fine-grained surgical phase recognition.
• To introduce a novel framework, SurgPETL, for Parameter-Efficient Image-to-Surgical-Video Transfer Learning.
• To evaluate the effectiveness and generalizability of this approach in complex surgical scenarios.

Main Methods:

• Developed SurgPETL, a parameter-efficient transfer learning framework for surgical phase recognition.
• Integrated an Adaptive Spatiotemporal Representation Modulation (ASRM) module with spatial and temporal adapters.
• Conducted extensive experiments using Vision Transformers (ViTs) pre-trained on diverse datasets.

Main Results:

• SurgPETL with ASRM demonstrated significant effectiveness on three challenging surgical datasets.
• The proposed framework outperformed existing parameter-efficient methods and state-of-the-art approaches.
• Achieved robust spatiotemporal modeling, capturing detailed spatial features and temporal connections.

Conclusions:

• SurgPETL-ASRM provides an efficient and effective solution for surgical phase recognition.
• The framework generalizes well across various surgical procedures, addressing data limitations.
• This approach maintains parameter efficiency and minimizes computational overhead in medical video analysis.