A framework for breast cancer classification using Multi-DCNNs.

Area of Science:

• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning (DL) and deep convolutional neural networks (DCNNs) are pivotal in modern image analysis.
• DCNN architectures like AlexNet, GoogleNet, and ResNet are widely used.
• Computer-aided diagnosis (CAD) systems aid radiologists in medical image interpretation.

Purpose of the Study:

• To develop and evaluate a novel CAD system for classifying breast cancer lesions in mammograms using DL.
• To compare different DL approaches for optimal breast cancer detection performance.
• To assess the impact of feature extraction, fusion, and dimensionality reduction on CAD system accuracy.

Main Methods:

• Four experiments were conducted: end-to-end DCNNs, DCNN features with Support Vector Machine (SVM) classifiers, deep feature fusion, and Principal Component Analysis (PCA) for dimensionality reduction.
• The system was evaluated on two public mammography datasets: CBIS-DDSM and MIAS.
• Deep features were extracted using DCNNs and classified using SVM with various kernels.

Main Results:

• Deep feature fusion achieved the highest accuracy among all tested methods, surpassing state-of-the-art CAD systems.
• Applying PCA to fused features did not improve accuracy but reduced computational cost and execution time.
• The study demonstrated the effectiveness of DL-based feature fusion for enhanced mammogram analysis.

Conclusions:

• Deep learning-based feature fusion offers a promising approach for improving the accuracy of breast cancer detection in mammography.
• While PCA can reduce computational load, it may not enhance diagnostic accuracy when applied to deep feature fusion.
• The developed CAD system provides a valuable tool to assist radiologists in breast cancer diagnosis.