Using Convolutional Neural Network with Cheat Sheet and Data Augmentation to Detect Breast Cancer in Mammograms

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Early breast cancer detection is crucial for reducing mortality.
• Deep learning, specifically Convolutional Neural Networks (CNNs), shows promise in accurate mammogram classification.
• A scarcity of labeled mammogram data hinders the training of effective CNN models.

Purpose of the Study:

• To propose a novel procedure to improve the accuracy of CNN-based mammogram classification.
• To address the challenge of limited labeled mammograms in training deep learning models.
• To aid imaging specialists in the accurate detection of normal and abnormal mammograms.

Main Methods:

• Developed a novel procedure incorporating a "cheat sheet" of classical attributes and data augmentation.
• The "cheat sheet" encodes artificial patterns to guide the CNN.
• Utilized data augmentation to increase the number of labeled mammograms for training.
• Conducted fifteen runs on four modified datasets from the MIAS dataset.

Main Results:

• The proposed procedure enhanced CNN accuracy by at least 12.2% and precision by at least 2.2.
• Achieved a mean accuracy of 92.1%, sensitivity of 91.4%, and specificity of 96.8%.
• The average area under the ROC curve was 94.9%.

Conclusions:

• The combination of a "cheat sheet" and data augmentation significantly improves CNN performance for mammogram analysis.
• This novel approach offers a viable solution for training accurate deep learning models with limited labeled data.
• The method aids in the reliable classification of mammograms, supporting early breast cancer detection.