An image retrieval framework for real-time endoscopic image retargeting.

Area of Science:

• Medical Imaging
• Computer Vision
• Gastroenterology

Background:

• Serial endoscopic examinations are crucial for early gastrointestinal malignancy detection.
• Retargeting for optical biopsy is difficult due to significant inter-examination tissue variations.
• Existing methods struggle with tissue changes and real-time retargeting requirements.

Purpose of the Study:

• To develop an image retrieval framework for robust inter-examination retargeting in endoscopic procedures.
• To create an image descriptor tolerant to long-term tissue variations.
• To implement a real-time descriptor matching method for efficient retargeting.

Main Methods:

• A novel image descriptor based on multi-scale regional intensity comparisons was developed for stability and discrimination.
• A real-time matching method using random forests to learn a hashing function for fast image comparison via Hamming distance was proposed.
• The framework was validated on a dataset of 13 in vivo gastrointestinal videos from six patients with significant time intervals between examinations.

Main Results:

• The proposed descriptor demonstrated superior retargeting precision-recall compared to alternative descriptors.
• The novel hashing method outperformed existing hashing approaches in speed and accuracy.
• The system achieved state-of-the-art retargeting performance on in vivo endoscopic videos.

Conclusions:

• A novel framework for optical biopsy in serial endoscopic examinations was successfully developed.
• The combination of the new descriptor and hashing method provides effective, state-of-the-art retargeting.
• The real-time capabilities allow for seamless integration into clinical workflows without disruption.