Venous thrombosis supervised image indexing and fuzzy retrieval

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Ultrasound Diagnostics

Background:

• Clinical assessment of venous thrombosis (VT) is crucial for risk evaluation.
• Current diagnostic methods like echogenicity analysis are operator-dependent and uncertain.
• Indexing and retrieving VT images for data mining and clinical support is complex.

Purpose of the Study:

• To propose a new approach for indexing and retrieving VT ultrasound images.
• To combine neural network characterization with fuzzy similarity for improved image analysis.
• To address the uncertainty and operator dependency in VT image diagnostics.

Main Methods:

• Utilized three types of image descriptors: sliding window, wavelet coefficients energy, and co-occurrence matrix.
• Processed descriptors using three different neural networks for VT characterization.
• Projected characterization values onto fuzzy membership functions and compared using fuzzy similarity.

Main Results:

• Neural networks produced equivalent VT characterizations from different image descriptors.
• Fuzzy similarity demonstrated increased image retrieval precision compared to nominal and Euclidean distances.
• The approach accounts for characterization uncertainty and allows user-defined feature prioritization.

Conclusions:

• The proposed neural network and fuzzy similarity approach is suitable for VT ultrasound image indexing and retrieval.
• This method enhances the precision of medical image retrieval, going beyond simple diagnostic class identification.
• It offers a more robust solution for managing and analyzing VT image data in clinical and research settings.