Mammographic image classification with deep fusion learning.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Mammographic image analysis is crucial for early breast cancer detection.
• Accurate classification of normal versus tumor tissues remains a challenge.
• Deep learning offers potential for improving mammogram interpretation.

Purpose of the Study:

• To develop and evaluate a deep fusion learning framework for mammographic image classification.
• To identify discriminative patterns distinguishing normal and tumor categories.
• To enhance the accuracy of abnormality recognition in mammograms.

Main Methods:

• A two-step deep fusion learning framework was designed using pre-trained models.
• Regions of Interest (ROIs) were extracted, and deep features were fused.
• Two models were proposed: Model1 for direct feature fusion and Model2 for cross-channel feature integration.
• Majority voting was employed for final ROI prediction.

Main Results:

• Model1 achieved 0.8906 overall accuracy, 0.913 recall, and 0.8077 precision for the tumor class.
• Model2 achieved 0.875 overall accuracy, 0.9565 recall, and 0.7586 precision for the tumor class.
• Both models demonstrated effectiveness in classifying mammographic abnormalities.

Conclusions:

• The proposed deep fusion learning framework shows promise for accurate mammographic image classification.
• The approach can aid in the recognition of abnormalities, potentially improving diagnostic capabilities.
• The study provides open-source code to facilitate further research in the field.