A Multi-Task Convolutional Neural Network for Semantic Segmentation and Event Detection in Laparoscopic Surgery.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Surgical Technology

Background:

• Intraoperative bleeding significantly complicates laparoscopic surgery, impairing surgical site visibility and requiring prompt management.
• Effective real-time monitoring of bleeding and surgical instruments is crucial for patient safety and surgical outcomes.

Purpose of the Study:

• To develop and evaluate a multi-task, end-to-end deep learning model for simultaneous blood accumulation detection and tool semantic segmentation in laparoscopic videos.
• To assess the model's performance using only RGB images without additional data or preprocessing.

Main Methods:

• A novel multi-task deep learning architecture was designed, utilizing a U-Net encoder as a shared backbone with two distinct branches for event classification and segmentation.
• The convolutional neural network was trained and tested on laparoscopic surgical videos, focusing on real-time processing capabilities.

Main Results:

• The model achieved a Dice Score of 81.89% for semantic segmentation of surgical tools and 90.63% accuracy for blood accumulation event detection.
• The shared backbone effectively leveraged features for both tasks, demonstrating the benefit of concurrent learning.
• Performance was satisfactory despite training solely on RGB images, indicating model robustness.

Conclusions:

• The developed multi-task deep learning model offers a promising approach for real-time video understanding in laparoscopic surgery.
• Concurrent blood detection and tool segmentation can be effectively achieved, potentially enhancing surgical safety and efficiency.
• This preliminary work lays the foundation for advanced AI-driven surgical assistance systems.