Interactive Surgical Training in Neuroendoscopy: Real-Time Anatomical Feature Localization Using Natural Language...

Area of Science:

• Medical Education Technology
• Computer Vision in Surgery
• Neurosurgery Training

Background:

• Surgical education, especially in neuroendoscopy, faces challenges balancing workflow optimization with trainee participation.
• Current training methods may not fully leverage interactive learning opportunities in complex procedures.

Purpose of the Study:

• To develop and evaluate a framework for accurate anatomical structure identification in neuroendoscopy using image and language data.
• To enhance interactive learning experiences and support trainee skill development in neuroendoscopy.

Main Methods:

• Utilized a transformer-based encoder-decoder architecture for multimodal input processing (images and language).
• Curated a dataset from recorded endoscopic third ventriculostomy (ETV) procedures for training and evaluation.
• Employed metrics such as R@n, IoU=θ, mIoU, and top-1 accuracy for benchmarking.

Main Results:

• Achieved 93.67% accuracy and 76.08% mean Intersection over Union (mIoU) on unseen data, outperforming existing methods.
• Demonstrated superior computational speed compared to other methodologies.
• Qualitative results confirmed precise localization of anatomical features.

Conclusions:

• The framework effectively localizes anatomical features using language descriptions, proving valuable for interactive clinical learning systems.
• This technology has the potential to significantly enhance surgical training in neuroendoscopy, leading to improved trainee skills and patient outcomes.