A Novel Algorithm for Breast Mass Classification in Digital Mammography Based on Feature Fusion

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Breast Cancer Diagnostics

Background:

• Accurate classification of breast masses is crucial for effective breast cancer screening.
• Current deep convolutional neural networks (CNNs) struggle with local-invariant features, limiting their performance with variations in imaging angles and geometric transformations.
• This limitation hinders the reliability of computer-aided diagnosis systems in mammography.

Purpose of the Study:

• To develop a novel model for classifying benign and malignant breast masses.
• To address the limitations of CNNs in handling local-invariant features for mammography mass classification.
• To improve the precision of breast mass screening through an integrated computer-aided diagnosis system.

Main Methods:

• A new model was proposed, integrating texton representation with deep CNN representation for mass classification.
• Rotation-invariant features from a maximum response filter bank were incorporated into the CNN-based classification.
• A fusion approach after implementing a reduction method was used to enhance CNN feature extraction capabilities.

Main Results:

• The proposed model was evaluated on public datasets: CBIS-DDSM, mini-MIAS, and INbreast.
• On the CBIS-DDSM dataset, the fusion model achieved superior performance with an area under the receiver operating curve of 0.97, accuracy of 94.30%, and specificity of 97.19%.
• The results demonstrate the model's effectiveness in overcoming CNN deficiencies in feature extraction.

Conclusions:

• The proposed novel model effectively integrates texton and deep CNN representations with rotation-invariant features.
• This approach significantly enhances the classification of breast masses, outperforming existing methods.
• The developed method holds promise for integration into computer-aided diagnosis systems for precise breast mass screening.