Gauze Detection and Segmentation in Minimally Invasive Surgery Video Using Convolutional Neural Networks.

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Robotics

Background:

• Automated medical instrument detection in laparoscopic video is crucial for surgical robotics and skill assessment.
• Research has overlooked surgical gauze detection due to the absence of annotated datasets.
• Gauze information is valuable for operating room tasks but remains underutilized.

Purpose of the Study:

• To introduce a novel annotated dataset for surgical gauze segmentation in laparoscopic videos.
• To evaluate baseline methods for gauze detection and segmentation.
• To demonstrate the feasibility of real-time, accurate surgical gauze segmentation.

Main Methods:

• Creation of a dataset with 4003 hand-labelled laparoscopic video frames.
• Implementation and analysis of baseline models: YOLOv3 for detection, coarse segmentation, and U-Net for segmentation.
• Performance evaluation based on detection accuracy, segmentation quality (IoU), and inference speed (FPS).

Main Results:

• YOLOv3 achieved real-time performance but with limited recall.
• Coarse segmentation provided acceptable results but lacked sufficient inference speed.
• The U-Net model demonstrated a balance between speed and quality, running over 30 FPS with an IoU of 0.85.

Conclusions:

• The proposed dataset enables research into surgical gauze segmentation.
• Convolutional neural networks, particularly U-Net, can achieve precise and real-time gauze segmentation.
• This advancement has implications for improving surgical workflow and robotic assistance.