RSDNet: Learning to Predict Remaining Surgery Duration from Laparoscopic Videos Without Manual Annotations

Area of Science:

• Computer Vision
• Medical Informatics
• Artificial Intelligence

Background:

• Accurate surgery duration estimation is crucial for operating room (OR) planning, impacting patient safety and resource optimization.
• Preoperative prediction of surgery duration is challenging due to patient variability, surgeon expertise, and intraoperative complexities.
• Current methods for remaining surgery duration (RSD) prediction often rely on time-consuming and expensive manual annotations.

Purpose of the Study:

• To develop a deep learning pipeline (RSDNet) for automatic intraoperative estimation of remaining surgery duration (RSD) using only laparoscopic video data.
• To create a scalable RSD prediction method that does not require manual annotation during training.
• To demonstrate the generalizability and performance of RSDNet across different surgical procedures.

Main Methods:

• Proposed RSDNet, a deep learning pipeline utilizing visual information from laparoscopic videos for intraoperative RSD estimation.
• Trained the model without manual annotations, enhancing scalability for diverse surgical types.
• Validated the approach on two large datasets: 120 cholecystectomy and 170 gastric bypass videos.

Main Results:

• RSDNet demonstrated significant outperformance compared to traditional methods for RSD estimation that do not use manual annotation.
• The model's generalizability was confirmed through testing on distinct surgical procedures.
• Feature visualization and interpretation provided insights into the learned representations within the deep learning network.

Conclusions:

• RSDNet offers an effective, annotation-free deep learning approach for intraoperative remaining surgery duration estimation.
• The method enhances scalability and accuracy in surgical planning, potentially improving patient care and resource management.
• The findings highlight the potential of visual data and deep learning for advancing surgical workflow optimization.