Breast cancer detection using deep convolutional neural networks and support vector machines

Area of Science:

• Medical Imaging and Diagnostics
• Artificial Intelligence in Healthcare
• Oncology

Background:

• Early breast cancer detection is crucial for improved patient outcomes.
• Existing computer-aided detection (CAD) systems require enhancement for higher accuracy in classifying malignant and benign breast tumors.
• Deep learning models offer potential for advanced feature extraction in mammography analysis.

Purpose of the Study:

• To develop and evaluate a novel CAD system for classifying breast mass tumors in mammography images.
• To leverage deep convolutional neural networks (DCNNs) and Support Vector Machine (SVM) classifiers for improved diagnostic accuracy.
• To compare the performance of different segmentation techniques and data augmentation strategies.

Main Methods:

• A computer-aided detection (CAD) system was designed incorporating two segmentation approaches: manual region of interest (ROI) selection and automated threshold/region-based segmentation.
• A Deep Convolutional Neural Network (DCNN), specifically AlexNet, was fine-tuned for feature extraction and classification.
• The fine-tuned DCNN was integrated with a Support Vector Machine (SVM) classifier. Data augmentation, including rotation, was employed to address limited dataset sizes.

Main Results:

• The DCNN achieved 71.01% accuracy with manually cropped ROIs.
• The highest Area Under the Curve (AUC) reached 0.88 (88%) using samples from both segmentation techniques.
• Utilizing the CBIS-DDSM dataset, the DCNN achieved 73.6% accuracy, leading to an SVM accuracy of 87.2% and an AUC of 0.94 (94%).

Conclusions:

• The proposed CAD system, integrating DCNN feature extraction with SVM classification, demonstrates significant potential for accurate breast cancer detection.
• The combined approach, particularly with the CBIS-DDSM dataset and automated segmentation, achieved superior performance compared to previous studies.
• This methodology offers a promising advancement in automated breast cancer diagnosis, highlighting the effectiveness of deep learning and data augmentation.