Developing an artificial intelligence model for phase recognition in robot-assisted radical prostatectomy.

Area of Science:

• Robotics
• Artificial Intelligence
• Surgical Technology

Background:

• Robot-assisted radical prostatectomy (RARP) is a complex procedure requiring precise surgical skill.
• Automating surgical phase recognition can aid in training and real-time feedback.
• Developing interpretable AI models is crucial for clinical adoption.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) for recognizing surgical phases in RARP.
• To assess the model's interpretability and cross-platform generalisability.
• To enhance trust and integration of AI in robotic surgery.

Main Methods:

• A CNN model (EfficientNet B7) was trained on 75 RARP cases using the hinotori robotic system.
• Seven distinct surgical phases were annotated across 808,774 video frames.
• Cross-platform validation was performed on 25 RARP cases using the da Vinci robotic system; interpretability was assessed using Gradient-weighted Class Activation Mapping (Grad-CAM).

Main Results:

• The CNN achieved 0.90 accuracy on the hinotori dataset but decreased to 0.64 on the da Vinci dataset, indicating cross-platform limitations.
• Phase-specific F1 scores ranged from 0.77 to 0.97, with lower performance in seminal vesicle and apical dissection phases.
• Grad-CAM visualizations highlighted the model's focus on anatomical structures, improving interpretability.

Conclusions:

• The CNN model shows high accuracy on a single robotic platform but requires further development for cross-platform consistency.
• Interpretability techniques are vital for building clinical trust and facilitating the integration of AI into surgical workflows.
• Continued refinement of AI models can advance the application of robotic surgery.