Improving the Generalisability of Deep CNNs by Combining Multi-stage Features for Surgical Tool Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Surgical Workflow Analysis

Background:

• Convolutional Neural Networks (CNNs) are widely used for surgical workflow analysis.
• Existing CNN models often suffer from limited generalization due to small, single-site surgical datasets.
• Assessing CNN performance on multi-institutional data is crucial but underexplored.

Purpose of the Study:

• To develop a novel CNN model that combines multi-stage features for accurate and generalized surgical tool classification.
• To evaluate the proposed model's generalization capability on diverse, multi-institutional surgical datasets.
• To improve the robustness of AI-driven assistive systems in surgery.

Main Methods:

• A CNN model integrating features from multiple stages was designed.
• The proposed model was extensively evaluated on three distinct surgical datasets (Cholec80, Cholec20, Gyna05).
• Performance was compared against baseline CNN models using mean Average Precision (mAP).

Main Results:

• The proposed CNN approach demonstrated superior generalization performance compared to baseline models.
• The model achieved mAP scores of 91.46% (Cholec80), 69.02% (Cholec20), and 37.14% (Gyna05).
• Generalization performance showed an improvement of approximately 7% over baseline CNN models.

Conclusions:

• The developed CNN method significantly enhances generalization capability for surgical tool classification.
• This advancement contributes to building more reliable assistive systems for surgeons.
• Improved assistive systems have the potential to enhance patient care and surgical outcomes.