Simultaneous shape and camera-projector parameter estimation for 3D endoscopic system using CNN-based grid-oneshot...

Area of Science:

• Medical Imaging
• Computer Vision
• Endoscopic Technology

Background:

• Accurate in situ endoscopic diagnosis and treatment rely on precise polyp size measurement.
• Current 3D endoscopic systems using active stereo techniques show promise but face challenges.
• Existing methods struggle with feature extraction stability and limited reconstruction areas.

Purpose of the Study:

• To enhance the accuracy and coverage of 3D endoscopic reconstruction for polyp size measurement.
• To address limitations in feature extraction stability and pattern projection area in active stereo 3D endoscopy.

Main Methods:

• A learning-based technique utilizing convolutional neural networks (CNNs) for stable feature extraction.
• An extended bundle adjustment technique to integrate multiple shapes for a consistent, larger reconstructed region.

Main Results:

• The proposed CNN-based approach improves the stability of feature extraction from endoscopic images.
• The extended bundle adjustment effectively expands the reconstructed region by integrating multiple shapes.
• Experimental evaluations demonstrate the superiority of the proposed techniques over previous methods.

Conclusions:

• The developed techniques significantly enhance the quality and coverage of 3D endoscopic reconstructions.
• This advancement supports more effective in situ endoscopic diagnosis and treatment through improved polyp size measurement.
• The study presents a robust solution for overcoming key limitations in active stereo 3D endoscopy.