BSc Thesis: SurgiToolVision

SurgiToolVision focuses on designing and evaluating a computer vision system that can automatically recognize surgical instruments and their state. The core system uses YOLO-based object detection models to identify handheld tools such as scalpels, clamps, shears, and needle holders from images or live camera feeds.

Beyond tool detection, the project includes state analysis for hinged instruments (for example, open vs. closed clamp or needle holder) and integrates the full machine learning pipeline: dataset preparation, model training, evaluation, and a user interface for image and live-camera inference.

Research Focus

A key contribution is an active vision approach for surgical tool recognition. Instead of relying only on a single static image, the system can leverage multiple viewpoints or request additional views when confidence is low. This helps in challenging situations such as partial occlusion or ambiguous tool appearances from specific angles.

Evaluation Goals

Through experiments across models, viewpoints, and system configurations, SurgiToolVision evaluates how multi-view information and instrument-state recognition improve reliability. The long-term motivation is to support future operating room technologies, including surgical workflow analysis, robotic assistance, and automated instrument tracking.