Clinical Report: Utilizing Advanced Deep Learning Techniques for Breast Cancer Identification
Overview
This study presents a two-step deep learning method for breast cancer detection in mammograms, particularly suited for resource-limited settings like Palestine. The proposed framework demonstrates results in tumor segmentation and classification.
Background
Breast cancer is the leading cancer among women in Palestine, accounting for over 34% of all cancer cases and 12% of cancer-related deaths. The healthcare system in Palestine faces significant challenges, including a lack of advanced diagnostic equipment.
Data Highlights
Model
Mean IoU
Dice Coefficient
Precision
Recall
Accuracy
AUC
U-Net (segmentation)
0.70
0.74
0.78
0.71
N/A
N/A
VGG16 (classification)
N/A
N/A
0.91
0.95 (malignant class)
91%
0.97
ResNet50
N/A
N/A
N/A
N/A
85%
N/A
MobileNet
N/A
N/A
N/A
N/A
82%
N/A
Key Findings
The U-Net segmentation model achieved a mean IoU of 0.70 and a Dice coefficient of 0.74.
The VGG16 classifier demonstrated 91% accuracy and an AUC of 0.97 on the Palestine evaluation subset.
Precision and recall for the malignant class were 0.91 and 0.95, respectively.
Comparison with ResNet50 and MobileNet was included.
The framework is designed to be scalable for low-resource environments.
Clinical Implications
Validation on larger datasets is necessary before clinical implementation.
Conclusion
This study provides a proof-of-concept for a deep learning pipeline for breast cancer detection in underserved areas.
