Abstract

This thesis explores the application of deep learning algorithms within extended reality to enhance preoperative planning in minimally invasive video-assisted thoracic surgery (VATS). VATS faced technical challenges, such as a limited field of view and complex anatomical structures, which require precise, patient-specific 3D modeling and intuitive data interaction for effective planning.
While deep learning, particularly U-shaped architectures, has emerged as a powerful approach for generating these models through automated segmentation of preoperative medical images, the growing number of U-shaped models with diverse network configurations and attention mechanisms requires systematic evaluation.
Our first contribution addresses this need through a comprehensive benchmark study of U-shaped models, focusing on their segmentation accuracy and computational complexity. The study reveals the effectiveness of CNN-based U-shaped architectures for thoracic anatomical segmentation, with residual blocks playing a crucial role in enhancing network performance. These findings provide essential guidance for model selection and development in surgical planning applications, where the balance between accuracy and computational efficiency is important.
Building upon these segmentation capabilities, our second contribution introduces an innovative extended reality system for optimizing trocar placement in VATS procedures. Optimal trocar placement is crucial to ensuring comprehensive thoracic cavity access, maintaining panoramic endoscopic visualization, and preventing instrument crowding. Our system features tailored visualization and interaction designs that enable surgeons to explore trocar configurations preoperatively using patient-specific 3D models. Preliminary evaluation demonstrates the system's efficiency, robustness, and user-friendliness, establishing its potential for clinical implementation while offering valuable insights for future surgical XR system development.