Content based medical image retrieval based on new efficient local neighborhood wavelet feature descriptor

Area of Science:

• Medical Imaging
• Computer Vision
• Signal Processing

Background:

• Effective content-based medical image retrieval (CBMIR) is crucial for medical diagnosis.
• Existing wavelet transform methods for CBMIR face challenges with high-dimensional feature vectors and limited performance.
• The Triplet Half-Band Filter Bank (THFB) offers enhanced wavelet filter properties.

Purpose of the Study:

• To introduce a new Local Neighborhood based Wavelet Feature Descriptor (LNWFD) for enhanced CBMIR.
• To address the limitations of existing methods in terms of feature vector length and retrieval accuracy.
• To develop a descriptor robust against illumination variations.

Main Methods:

• Utilized Triplet Half-Band Filter Bank (THFB) for single-level wavelet decomposition into four sub-bands.
• Exploited relationships among wavelet coefficients within a 3x3 neighborhood window in each sub-band to form the LNWFD.
• Employed Manhattan distance for computing similarity between query and database feature vectors.

Main Results:

• The LNWFD descriptor demonstrated robustness against illumination.
• Achieved high average retrieval precisions: 71.45% (OASIS-MRI, top 10), 99.51% (NEMA-CT, top 10), and 55.51% (Emphysema-CT, top 50).
• Experimental results confirmed superior performance compared to existing descriptors.

Conclusions:

• The proposed LNWFD offers a significant advancement in CBMIR.
• The descriptor effectively captures local information within wavelet sub-bands, leading to improved retrieval performance.
• LNWFD provides a robust and efficient solution for retrieving images from large medical databases.