HED-Net: a hybrid ensemble deep learning framework for breast ultrasound image classification

Breast cancer is a life-threatening disease that affects millions of women worldwide. Diagnosing breast cancer through imaging tests like ultrasound can be challenging, especially for radiologists. The current methods often rely on human expertise and can be time-consuming, which may lead to delays in diagnosis and treatment. Therefore, developing an efficient and accurate method for breast ultrasound image classification is crucial.

To tackle this challenge, researchers proposed a hybrid ensemble deep learning framework called HED-Net. This framework combines the strengths of three different convolutional neural network models: EfficientNetB7, DenseNet121, and ConvNeXtTiny. Each model is trained independently on breast ultrasound image datasets to capture complementary representations.

The HED-Net framework integrates the features extracted by each model into a high-dimensional unified representation. This representation is then used to learn non-linear decision boundaries using XGBoost as the feature fusion classifier. Additionally, a soft voting ensemble method averages the predicted probabilities of individual models to improve overall performance.

The HED-Net framework was evaluated on three breast ultrasound image datasets: BUSI, BUS-UCLM, and UDIAT. The results showed significant improvements in accuracy, precision, recall, F1 score, and AUC values compared to individual models.

The HED-Net framework has the potential to revolutionize breast ultrasound image classification by enhancing diagnostic accuracy and reducing interpretation time. This can be particularly beneficial in resource-limited environments where expert radiologists are scarce. The findings of this study highlight the importance of developing innovative AI solutions for healthcare applications.