Dual-discriminator network-based classification method for breast ultrasound imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer poses a significant global health challenge, necessitating early detection through ultrasound imaging.
• Challenges in medical image classification include small datasets, data imbalance, and the need for robust feature extraction and interpretability.
• Transfer learning from natural images is common but faces limitations with medical data.

Purpose of the Study:

• To address data imbalance in breast ultrasound image classification.
• To enhance the accuracy and interpretability of breast cancer diagnosis using AI.
• To improve early detection and patient outcomes for breast cancer.

Main Methods:

• Developed a novel Dual-Discriminator Generative Adversarial Network (GAN) for iterative data synthesis on unbalanced datasets.
• Incorporated channel and spatial attention mechanisms for intricate feature recognition in classification.
• Validated the approach on the Breast Ultrasound Images Dataset (BUSI) and a self-constructed dataset.

Main Results:

• Achieved a top accuracy of 96.0% on the BUSI dataset.
• Attained 95.8% accuracy on the self-constructed evaluation dataset.
• Demonstrated competitive performance against state-of-the-art methods.

Conclusions:

• The proposed method effectively addresses data imbalance and enhances classification accuracy.
• Attention mechanisms improve the recognition of critical diagnostic features in ultrasound images.
• Visualization methods confirm the practical diagnostic value for breast cancer identification.