Similarity searching for chest CT images based on object features and spatial relation maps

Area of Science:

• Medical Imaging
• Computer Science
• Radiology

Background:

• Content-based image retrieval (CBIR) is crucial for managing large medical image databases.
• Accurate segmentation of anatomical structures in chest CT scans is a prerequisite for effective retrieval.
• Existing retrieval methods may lack specificity for complex medical imaging data.

Purpose of the Study:

• To develop an object-based image retrieval system for chest CT image databases.
• To enhance the accuracy and efficiency of searching within medical imaging archives.
• To leverage anatomical knowledge for improved image segmentation and feature extraction.

Main Methods:

• An image segmentation method combining chest anatomical knowledge and watershed algorithm to identify mediastinum and lung lobes.
• Utilizing attributed relational graphs (ARGs) to describe segmented object features.
• Constructing an image database using feature vectors and implementing "query by example" and "query by object" search modes with Euclidean distance for similarity measurement.

Main Results:

• The proposed system achieved an average precision of approximately 80% in retrieving chest CT images.
• Object-focused queries demonstrated superior performance compared to standard "query by example" methods.
• The system successfully outputs the 30 most similar images based on query criteria.

Conclusions:

• The developed object-based image retrieval system offers a highly effective and automated solution for chest CT databases.
• Integrating anatomical knowledge into segmentation significantly improves retrieval accuracy.
• Further research into object feature-based querying can optimize medical image retrieval performance.