High accuracy breast cancer classification with BIRADS and coclustering

Area of Science:

• Medical Imaging
• Machine Learning
• Oncology

Background:

• Breast cancer classification relies on complex multi-phase methods (segmentation, feature extraction, classification).
• Low-level image features pose challenges for clinical interpretation by physicians.
• Imbalanced datasets (malignant vs. benign cases) hinder accurate classification model training.

Purpose of the Study:

• To develop a novel breast cancer classification method utilizing high-level Breast Imaging Reporting and Data System (BI-RADS) features.
• To overcome limitations of existing methods, including interpretability and data imbalance.
• To enhance diagnostic accuracy and clinical utility in breast cancer detection.

Main Methods:

• An improved Synthetic Minority Oversampling Technique (SMOTE) was employed to address data imbalance by generating synthetic minority class samples.
• Co-clustering was utilized to mine diagnostic rules from the data.
• Adaboost algorithm was applied to construct a strong classifier based on the mined rules.

Main Results:

• The proposed method demonstrated significant improvements in accuracy, precision, recall, and F1-score (over 5%) compared to existing methods on two public datasets.
• The method maintained superior accuracy (over 5% higher) even under varying degrees of data imbalance.
• The use of high-level BI-RADS features enhanced the interpretability of classification outcomes.

Conclusions:

• The novel breast cancer classification approach based on high-level BI-RADS features effectively handles data imbalance and improves diagnostic performance.
• This method offers a more interpretable and accurate alternative to traditional low-level feature-based classification techniques.
• The findings suggest a promising direction for improving automated breast cancer diagnosis in clinical settings.