Revolutionizing breast ultrasound diagnostics with EfficientNet-B7 and Explainable AI

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is a major global health concern, requiring accurate diagnostic tools.
• Current Convolutional Neural Network (CNN) methods face challenges with imbalanced datasets and subtle image variations in breast ultrasound classification.
• This limits diagnostic accuracy, particularly for malignant tumors.

Purpose of the Study:

• To develop a robust and interpretable breast ultrasound image classification system.
• To improve diagnostic accuracy by addressing class imbalances and enhancing model robustness.
• To increase the reliability and clinical adoption of automated breast cancer detection systems.

Main Methods:

• Fine-tuning the EfficientNet-B7 architecture on the Breast Ultrasound Images (BUSI) dataset.
• Employing data augmentation techniques (RandomHorizontalFlip, RandomRotation, ColorJitter) to improve minority class representation.
• Integrating Explainable AI (XAI) methods, specifically Grad-CAM, for model interpretability.

Main Results:

• Achieved a high classification accuracy of 99.14% for breast ultrasound images.
• Demonstrated superior performance compared to traditional CNN-based approaches.
• XAI techniques provided visual insights into classification decision-making processes.

Conclusions:

• The proposed EfficientNet-B7 framework with XAI offers a highly accurate and interpretable solution for breast cancer diagnosis.
• The methodology effectively handles class imbalances and enhances model robustness.
• This approach supports clinical decision-making and advances automated diagnostic capabilities for early breast cancer detection.