Generalization of Deep Learning Gesture Classification in Robotic-Assisted Surgical Data: From Dry Lab to...

Area of Science:

• Robotics
• Surgical Technology
• Machine Learning

Background:

• Robotic-assisted minimally invasive surgery (RAMIS) is prevalent, necessitating methods to evaluate surgeon performance.
• Classifying surgical gestures is crucial for characterizing surgeon performance in complex procedures.

Purpose of the Study:

• To assess the generalizability of gesture classification algorithms trained on dry-lab data to real surgical scenarios.
• To improve the accuracy of surgical gesture classification in clinical settings.

Main Methods:

• A Long Short-Term Memory (LSTM) network was trained for gesture classification using dry-lab and clinical-like datasets.
• The study evaluated the impact of rotation augmentation and the inclusion of patient-side manipulator (PSM) joint angles on classification performance.

Main Results:

• Networks trained solely on the JIGSAWS dry-lab dataset showed poor generalization to other dry-lab and clinical-like data.
• Rotation augmentation improved dry-lab performance but not clinical-like data performance.
• Adding PSM joint angle features and training on clinical data significantly enhanced classification accuracy for clinical scenarios.

Conclusions:

• The JIGSAWS dataset alone is insufficient for training robust clinical gesture classification models.
• Combining insights from dry-lab data for network architecture with a small amount of clinical data training yields acceptable performance.
• Improved gesture classification algorithms are vital for advancing understanding of surgeon control, skill evaluation, and surgical automation in RAMIS.