Investigating the detection of breast cancer with deep transfer learning using ResNet18 and ResNet34

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer incidence is rising in underdeveloped nations, particularly in Africa, due to delayed diagnosis.
• Early cancer detection is crucial for effective treatment and improved patient outcomes.
• Computer-aided diagnosis and medical image analysis can accelerate and automate cancer classification.

Purpose of the Study:

• To investigate the use of transfer learning with Residual Network 18 (ResNet18) and Residual Network 34 (ResNet34) architectures for breast cancer detection in mammography images.
• To develop a functional prototype application for radiologists to assist in breast cancer diagnosis.
• To evaluate the effectiveness of deep learning models in improving breast cancer detection accuracy, with a focus on African healthcare scenarios.

Main Methods:

• Utilized transfer learning techniques applying ResNet18 and ResNet34 deep learning architectures.
• Trained models on a dataset of 1,200 breast x-ray mammography images from the National Radiological Society (NRS) archives.
• Classified images into four categories: implant cancer negative, implant cancer positive, cancer negative, and cancer positive for multi-class analysis.

Main Results:

• Achieved 86.7% validation accuracy for binary classification (benign vs. malignant) using ResNet34.
• Achieved 92% validation accuracy for binary classification (benign vs. malignant) using ResNet18.
• Developed a prototype web application demonstrating the ResNet18 model's performance for clinical use.

Conclusions:

• Transfer learning significantly enhances the accuracy of breast cancer detection using deep learning models.
• The developed ResNet18 model shows promise for assisting medical professionals in early breast cancer diagnosis.
• This approach offers a valuable tool to improve cancer care, particularly in resource-limited settings like Africa.