A similarity measure method combining location feature for mammogram retrieval

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Machine Learning

Background:

• Breast cancer poses a significant mortality risk, underscoring the need for effective early detection methods.
• Mammogram retrieval systems are crucial for identifying early breast lesions and improving diagnostic accuracy.
• Current retrieval methods can be enhanced by incorporating specific feature sets for image similarity measures.

Purpose of the Study:

• To develop an advanced mammogram retrieval system utilizing a novel similarity measure.
• To improve the accuracy of identifying similar mammograms for early breast cancer detection.
• To integrate both location and content-based features for a more robust image retrieval process.

Main Methods:

• A similarity measure combining lesion location and content features for mammogram retrieval was proposed.
• Lesion segmentation and Coherent Point Drift registration were used to extract standard location features.
• Content features, including Histogram of Oriented Gradients and Local Binary Pattern, were extracted and analyzed using Earth Mover's Distance.
• Location and content similarities were fused to generate an overall image fusion similarity score.

Main Results:

• The experimental database comprised 440 mammograms from Chinese women.
• A fusion of 40% lesion location similarity and 60% content similarity yielded optimal results.
• The method achieved a precision of 0.83, recall of 0.76, and a comprehensive indicator of 0.79.
• High satisfaction (96.0%) and a low variance (17.7) were reported, indicating reliable performance.

Conclusions:

• The proposed mammogram retrieval method demonstrates significant advantages in precision and recall.
• This approach offers a substantial improvement over traditional content-based image retrieval techniques.
• The combined feature approach enhances the effectiveness of mammogram retrieval for early breast cancer detection.