BreastUS-Net: An Attention-Guided Dual-Branch Network With Feature Fusion for Fine-Grained Breast Tumor...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Deep learning (DL) and computer vision show promise in breast cancer (BC) detection via ultrasound, but challenges remain in clinical applicability due to complex pipelines and limited datasets.
• Existing methods often struggle with single-task focus and manual feature extraction, hindering widespread adoption in clinical settings.

Purpose of the Study:

• To develop a novel deep learning architecture, BreastUS-Net, for hierarchical breast cancer classification using diverse ultrasound datasets.
• To enhance feature extraction, representation, and diagnostic accuracy in breast ultrasound imaging through an integrated, multi-component model.

Main Methods:

• Proposed BreastUS-Net utilizes a dual-branch MobileNet architecture with fine-tuned and frozen layers for comprehensive feature capture.
• Feature fusion, aggregation, and refinement modules are employed, alongside multi-head self-attention (MHSA) and an orthogonal softmax layer (OSL) for improved accuracy and robustness.
• The model was trained and validated on six diverse datasets from multiple centers, including clinical and public datasets, and evaluated using explainable AI (XAI) techniques.

Main Results:

• BreastUS-Net achieved state-of-the-art performance, with accuracies of 94.48% on a clinical dataset and 94.23% on the BUSI dataset.
• The model effectively captures task-specific and general features, reduces complexity, highlights relevant patterns, and mitigates overfitting.
• MHSA and OSL integration enhanced diagnostic region highlighting, accuracy, and robustness, while XAI techniques improved trust in predictions.

Conclusions:

• BreastUS-Net demonstrates significant potential for improving breast cancer diagnosis and personalized treatment through accurate and robust ultrasound image analysis.
• The hierarchical classification approach and integrated architectural components offer a promising solution to the challenges in automated breast ultrasound analysis.
• The use of diverse datasets and XAI techniques underscores the model's clinical relevance and trustworthiness.