LoViT: Long Video Transformer for surgical phase recognition.

Area of Science:

• Computer Vision
• Medical Imaging
• Artificial Intelligence

Background:

• Surgical phase recognition is crucial for developing tools to quantify performance and oversee surgical workflows.
• Current methods struggle with frame-level supervision, leading to prediction errors and poor fusion of local/global features in long surgical videos.

Purpose of the Study:

• To develop an advanced method for surgical phase recognition that addresses limitations in existing approaches for long videos.
• To enhance the accuracy and robustness of surgical phase recognition by incorporating temporal dynamics and phase transitions.

Main Methods:

• Proposed Long Video Transformer (LoViT), a two-stage method featuring a temporally-rich spatial feature extractor and a phase transition map.
• Employed a multiscale temporal aggregator with cascaded L-Trans and G-Informer modules for processing temporal information.
• Utilized phase transition-aware supervision for surgical phase classification.

Main Results:

• LoViT demonstrated superior performance over state-of-the-art methods on the Cholec80 and AutoLaparo datasets.
• Achieved a 2.4 pp improvement in video-level accuracy on Cholec80 and 3.1 pp on AutoLaparo compared to Trans-SVNet.
• Consistently outperformed existing approaches across datasets with varying surgical procedures and temporal characteristics.

Conclusions:

• The proposed LoViT method effectively enhances surgical phase recognition by leveraging temporally-rich features and phase transition maps.
• LoViT achieves state-of-the-art performance, showcasing its potential for real-world surgical workflow analysis.
• The approach proves effective for diverse surgical procedures and video complexities.