Qualitative Comparison of Image Stitching Algorithms for Multi-Camera Systems in Laparoscopy.

Area of Science:

• Minimally Invasive Surgery
• Computer Vision
• Medical Imaging

Background:

• Multi-camera systems enhance the surgeon's field of view in laparoscopy.
• Current laparoscopic multi-camera systems use basic stitching algorithms.
• Advanced algorithms from other fields lack evaluation in surgical settings.

Purpose of the Study:

• To evaluate classical and state-of-the-art image stitching techniques for laparoscopy.
• To develop a versatile simulated environment and dataset for multi-view laparoscopic imaging.
• To identify challenges and propose improvements for laparoscopic image stitching.

Main Methods:

• Developed a simulated environment for generating multi-view laparoscopic image datasets.
• Evaluated global homography, classical, and deep learning-based image stitching techniques.
• Assessed algorithm performance on diverse simulated laparoscopic scenarios.

Main Results:

• Classical global homography techniques yield clinically unsatisfactory results.
• State-of-the-art algorithms show poor alignment and distortions in laparoscopic contexts.
• Unsupervised deep learning approaches demonstrated potential but require adaptation.

Conclusions:

• Existing image stitching algorithms, even advanced ones, are not directly suitable for laparoscopy.
• Laparoscopic conditions present unique challenges for image stitching.
• Further research is needed to adapt and improve stitching techniques for enhanced surgical visualization.