Addressing multi-label imbalance problem of surgical tool detection using CNN

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Technology

Background:

• Current surgical tool detection methods often overlook tool co-occurrence and class imbalance issues.
• Supervised one-vs-all or multi-class classification techniques are standard but have limitations.

Purpose of the Study:

• To propose a fully automated surgical tool detection framework for endoscopic video streams.
• To address limitations of existing methods by considering tool co-occurrence and class imbalance.

Main Methods:

• Formulating tool detection as a multi-label classification task, treating tool co-occurrences as distinct classes.
• Analyzing and addressing imbalance in tool co-occurrences using stratification techniques during convolutional neural network (CNN) training.
• Implementing temporal smoothing as an online post-processing step for improved prediction accuracy.

Main Results:

• Quantitative analysis on the M2CAI16 tool detection dataset demonstrated the effectiveness of the proposed framework.
• Stratification and temporal smoothing were shown to be crucial components for enhancing tool detection performance.
• The overall framework significantly improved surgical tool detection accuracy.

Conclusions:

• Empirical results confirm the need for specialized techniques to handle tool imbalance in surgical video analysis.
• The proposed framework demonstrates superiority over state-of-the-art methods for automated surgical tool detection.
• The findings highlight the potential of advanced machine learning techniques in improving surgical tool recognition.