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

Background: Breast cancer is a malignant tumor with significant morbidity and mortality worldwide. Early detection and diagnosis via ultrasound imaging are crucial for effective treatment and improved patient outcomes. Transferring learning from large natural image datasets to smaller medical datasets has been a common strategy to enhance model performance due to the scarcity of labeled medical images. However, there are several challenges in applying these methods to medical image classification, including the small dataset size, data imbalance, the need for strong feature extraction capabilities, and the necessity for model interpretability. This study aims to address data imbalance and enhance both the accuracy and interpretability of breast ultrasound image classification for breast cancer diagnosis. Methods: We introduce a novel Dual-Discriminator Generative Adversarial Networks (GAN) for iterative data synthesis on unbalanced datasets. We introduce channel and spatial attention mechanisms to recognize intricate details during classification. Extensive experiments on various benchmarks demonstrate that our method achieves competitive performance compared to state-of-the-art methods. Our method shows improved performance on both the Breast Ultrasound Images Dataset (BUSI) and our self-constructed dataset. Results: Our method achieved a top accuracy of 96.0% on the BUSI dataset and 95.8% on our self-constructed evaluation dataset. Conclusions: Qualitative evaluation through visualization methods further supports the practical diagnostic value of our method in identifying relevant image features for classification.