Intraoperative Detection of Surgical Gauze Using Deep Convolutional Neural Network

Area of Science:

• Medical Imaging
• Artificial Intelligence in Surgery
• Surgical Safety

Background:

• Surgical gauze is crucial for hemostasis during laparoscopic procedures.
• Retained surgical gauze poses significant risks, including the need for reoperation and increased surgical complications.

Purpose of the Study:

• To develop and evaluate a deep learning neural network model for detecting surgical gauze within laparoscopic surgical videos.
• To enable real-time tracking and recording of gauze presence to enhance surgical safety and prevent retained surgical items.

Main Methods:

• A deep learning model utilizing the YOLO (You Only Look Once)v5x6 architecture was trained on a dataset of surgical videos.
• The model's performance was evaluated on a separate testing group, calculating metrics such as Positive Predicted Value (PPV), sensitivity, and mean Average Precision (mAP).
• A timeline of gauze presence was generated by the model and compared against human annotations for accuracy assessment.

Main Results:

• The trained model achieved a PPV of 0.920, sensitivity of 0.828, and mAP of 0.881 in the testing group.
• The model demonstrated a fast inference time of 11.3 milliseconds per image.
• An average accuracy of 0.899 was achieved when incorporating marking and filtering processes.

Conclusions:

• Deep learning technology can successfully detect surgical gauze in real-time from surgical videos.
• The developed model offers rapid and accurate gauze detection, facilitating real-time tracing during laparoscopic surgery.
• This AI-driven approach has the potential to significantly improve surgical safety by assisting surgeons in accounting for all surgical materials.