Multi-Stage Temporal Convolutional Network with Moment Loss and Positional Encoding for Surgical Phase Recognition.

Area of Science:

• Medical image analysis
• Computer vision in surgery

Background:

• Surgical video analysis is crucial for medical applications.
• Accurate surgical phase recognition aids intraoperative and postoperative decision-making.

Purpose of the Study:

• To propose an efficient phase recognition network, MomentNet, for cholecystectomy endoscopic videos.
• To improve surgical phase prediction accuracy using novel techniques.

Main Methods:

• MomentNet utilizes a multi-stage temporal convolutional network architecture.
• A new loss function is introduced to constrain undesirable phase transitions and prevent over-segmentation.
• Positional encoding, label smoothing, and an improved backbone network are incorporated.

Main Results:

• MomentNet achieved 92.31% accuracy on the Cholec80 dataset for phase recognition.
• This represents a 4.55% improvement over the baseline architecture.
• The novel loss function and positional encoding significantly enhance prediction accuracy.

Conclusions:

• MomentNet demonstrates superior performance in surgical phase recognition for cholecystectomy.
• The proposed methods effectively address challenges like over-segmentation and temporal context.
• This advancement holds promise for improved surgical video analysis and clinical applications.