HBMD-Net: Feature Fusion Based Breast Cancer Classification with Class Imbalance Resolution

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is a leading global health threat, necessitating accurate diagnostic tools.
• Current computer-aided diagnosis systems often rely on deep learning (DL) but lack reliability due to insufficient feature integration.
• Limited and imbalanced breast cancer datasets pose significant challenges for model training and generalization.

Purpose of the Study:

• To develop a hybrid breast mass detection-network (HBMD-Net) for enhanced breast cancer classification.
• To address critical challenges of class imbalance and the need for combined local and global feature analysis.
• To improve the precision of tumor detection in breast ultrasound images.

Main Methods:

• Implemented the borderline synthetic minority over-sampling technique (BSMOTE) to mitigate class imbalance.
• Employed a feature fusion approach combining ResNet50 for global features and Histogram Orientation Gradient (HOG) for local features.
• Integrated Region of Interest (ROI) segmentation and a block matching and 3D (BM3D) denoising filter to enhance classification accuracy and reduce noise.

Main Results:

• The HBMD-Net achieved high classification accuracies of 99.14% on the BUSI dataset and 94.49% on the UDIAT dataset.
• The hybrid approach effectively combined local and global features, outperforming methods relying on a single feature type.
• The integration of BSMOTE, feature fusion, ROI segmentation, and BM3D denoising significantly improved diagnostic performance.

Conclusions:

• HBMD-Net offers a robust solution for breast cancer classification, effectively handling class imbalance and feature heterogeneity.
• The proposed model demonstrates superior performance in breast ultrasound image analysis, paving the way for more reliable computer-aided diagnosis.
• This hybrid approach highlights the importance of integrating diverse features and noise reduction techniques for accurate medical image classification.