Deep Learning Model for Real‑time Semantic Segmentation During Intraoperative Robotic Prostatectomy.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in surgery
• Robotic surgery

Background:

• Deep learning, particularly convolutional neural networks (CNNs), has advanced medical imaging analysis.
• Semantic segmentation is crucial for AI applications in surgery, including AI-assisted surgery, training, and skill assessment.
• This study investigates a CNN model for real-time segmentation during robot-assisted radical prostatectomy (RALP).

Purpose of the Study:

• To evaluate the performance of a CNN-based deep learning model for real-time semantic segmentation in RALP.
• To assess the model's accuracy in identifying surgical instruments and critical anatomical structures during the procedure.

Main Methods:

• Intraoperative videos from RALP procedures were utilized.
• A reinforcement U-Net model performed segmentation of instruments, bladder, prostate, and seminal vesicle-vas deferens.
• Performance was evaluated using Dice coefficient, intersection over union (IoU), and accuracy on a 7:2:1 split training, validation, and test dataset.

Main Results:

• The model achieved high mean Dice scores for instruments (0.96), prostate (0.85), and seminal vesicle-vas deferens (0.84), with a score of 0.74 for the bladder.
• Overall performance on test data showed a mean Dice coefficient of 0.85, IoU of 0.77, and accuracy of 0.85.
• Limitations included sample size, surgical method diversity, incomplete procedures, and lack of external validation.

Conclusions:

• CNN-based semantic segmentation accurately recognizes surgical instruments and anatomy in real-time during RALP.
• Deep learning models can identify surgical anatomy and offer potential for real-time guidance in robotic surgery.
• The study demonstrates the effectiveness of deep learning segmentation for enhancing robotic prostatectomy procedures.