Domain generalization improves end-to-end object detection for real-time surgical tool detection

Area of Science:

• Computer Vision
• Medical Imaging
• Surgical Robotics

Background:

• Real-time surgical tool detection is crucial for computer-assisted endoscopic surgery.
• Current multi-step detection methods suffer from performance limitations and complexity.
• Limited datasets in surgical training lead to poor generalization of deep learning models.

Purpose of the Study:

• To develop an end-to-end transformer-based architecture for surgical tool detection.
• To enhance model generalization across diverse surgical domains and datasets.
• To improve the real-time performance of surgical tool detection systems.

Main Methods:

• An end-to-end transformer architecture was applied for surgical tool detection.
• A latent feature space using variational encoding was incorporated to capture common intra-domain information.
• Rank constraints were used to model linear dependencies between domains, improving cross-domain generalization.

Main Results:

• The proposed method demonstrated improved performance on out-of-domain data, indicating enhanced generalization.
• Inference achieved up to 138 frames per second, offering a significant speedup over existing approaches.
• Experimental results on three datasets validated the method's effectiveness and domain generalization capabilities.

Conclusions:

• The transformer-based end-to-end approach offers a more efficient and accurate solution for surgical tool detection.
• The integration of a latent feature space effectively addresses the challenge of limited datasets and improves cross-domain generalization.
• This advancement holds promise for enhancing the capabilities of computer-assisted endoscopic surgery.