Semi-automated query construction for content-based endomicroscopy video retrieval.

Area of Science:

• Medical Imaging
• Computer Science

Background:

• Content-based video retrieval (CBVR) aids medical video interpretation, especially endomicroscopy.
• Query formulation for CBVR is challenging, impacting retrieval accuracy.
• Uncut videos as queries lead to performance degradation due to diverse tissue types.

Purpose of the Study:

• To develop a semi-automated methodology for efficient and relevant query creation in CBVR for medical videos.
• To enhance the reproducibility and consistency of medical video retrieval results.
• To improve physician's ability to query and interpret endomicroscopic videos.

Main Methods:

• A semi-automated methodology for physician-guided query construction.
• Indirect validation using per-video classification with histopathological ground-truth.
• Direct validation based on perceived inter-video visual similarity.

Main Results:

• The proposed method significantly outperforms using uncut videos as queries.
• The system's performance approaches that of manual query construction by experts.
• Computed inter-video similarity correlates significantly with expert-perceived similarity.

Conclusions:

• The semi-automated method offers a simple and efficient way to create meaningful queries for medical video retrieval.
• This approach enhances the reliability and consistency of CBVR systems in clinical settings.
• The system effectively captures expert-defined visual similarity for improved retrieval accuracy.