MT4MTL-KD: A Multi-Teacher Knowledge Distillation Framework for Triplet Recognition.

Area of Science:

• Computer Vision
• Medical Imaging
• Machine Learning

Background:

• Surgical triplet recognition is vital for analyzing surgical videos, involving instrument, verb, and target identification.
• Current methods struggle with imbalanced data and integrating local/global context for accurate triplet association.

Purpose of the Study:

• To develop a novel framework, MT4MTL-KD, to enhance multi-task learning for surgical triplet recognition.
• To address challenges of class imbalance and context modeling in surgical video analysis.

Main Methods:

• A multi-teacher knowledge distillation framework (MT4MTL-KD) was proposed, using teacher models trained on balanced sub-tasks to guide student learning.
• Different backbone architectures for teacher and student models were employed to integrate local and global context.
• A feature attention module (FAM) was introduced to align semantic knowledge between triplet tasks and sub-tasks.

Main Results:

• The MT4MTL-KD framework demonstrated superior performance on the CholecT45 dataset across 5-fold cross-validation and CholecTriplet challenge splits.
• Significant improvements of up to 6.4% were achieved on the cross-validation split compared to state-of-the-art methods.
• The proposed methods effectively mitigated issues related to imbalanced class distribution and improved context integration.

Conclusions:

• MT4MTL-KD offers a robust solution for surgical triplet recognition by effectively handling data imbalance and enhancing feature representation.
• The framework shows significant potential for advancing the practical applications of surgical video analysis and understanding.
• Future work could explore further refinements in attention mechanisms and knowledge distillation strategies for even greater accuracy.