Deep transfer learning for breast cancer detection in underserved regions.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Deep Learning for Cancer Detection
• Health Informatics in Low-Resource Settings

Background:

• Breast cancer is the leading cancer among women in Palestine, contributing to significant mortality.
• The Palestinian healthcare system faces challenges including equipment shortages and limited advanced diagnostics.
• There is a critical need for cost-effective and scalable diagnostic tools in underserved regions.

Purpose of the Study:

• To propose and evaluate a novel two-step deep learning framework for breast cancer detection in mammograms.
• To assess the feasibility of this framework in low-resource environments, specifically Palestine.
• To develop a diagnostic support tool capable of distinguishing between benign and malignant tumors.

Main Methods:

• A U-Net architecture with a VGG16 encoder was utilized for lesion segmentation on the CBIS-DDSM dataset.
• A VGG16 classification model, pre-trained on ImageNet, was employed to classify segmented regions as benign or malignant.
• The framework was evaluated using the CBIS-DDSM dataset for training and validation, and a Palestine Hospital dataset for external testing.

Main Results:

• The U-Net segmentation model achieved a Dice coefficient of 0.74 and a recall of 0.71.
• The VGG16 classifier demonstrated 91% accuracy, 0.95 recall for malignant cases, and an AUC of 0.97 on the Palestine evaluation subset.
• The proposed deep learning approach outperformed ResNet50 (85% accuracy) and MobileNet (82%) in classification tasks.

Conclusions:

• The developed deep learning pipeline is a promising proof-of-concept for breast cancer detection in resource-limited settings.
• The framework's feasibility in low-resource environments suggests potential for scalable and cost-effective diagnostic assistance.
• Further validation on larger, locally annotated datasets is recommended before clinical deployment.