Medical image categorization and retrieval for PACS using the GMM-KL framework

Area of Science:

• Medical Imaging
• Computer Vision
• Information Retrieval

Background:

• Content-based image retrieval (CBIR) enhances radiologist search capabilities by integrating visual analysis with text-based search.
• Current CBIR systems are being integrated into picture archiving and communication systems.
• Medical image archives require robust categorization for efficient retrieval.

Purpose of the Study:

• To develop an image representation and matching framework for automated medical image categorization.
• To improve content-based image retrieval (CBIR) in medical image archives.
• To enable automatic determination of body region and imaging modality from image content.

Main Methods:

• A probabilistic image representation using Gaussian mixture modeling (GMM).
• Information-theoretic image matching utilizing the Kullback-Leibler (KL) measure.
• Multidimensional feature space incorporating intensity, texture, and spatial information, with unsupervised clustering for region extraction.

Main Results:

• Achieved a 97.5% classification rate for categorizing X-ray images by body region.
• Demonstrated favorable comparison with existing global and local representation schemes.
• Precision-recall curves indicated strong retrieval performance against state-of-the-art techniques.

Conclusions:

• The GMM-KL framework provides an effective solution for medical image categorization and retrieval.
• The illumination-invariant representation successfully handles challenges posed by poor contrast and intensity variations in radiological images.
• The framework enables learning of category models for image comparison, enhancing search capabilities in medical archives.