Efficient Ultrasound Breast Cancer Detection with DMFormer: A Dynamic Multiscale Fusion Transformer.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Ultrasound imaging presents challenges in breast cancer screening due to noise, blur, and complex tissue structures.
• Accurate differentiation between benign and malignant masses is crucial for effective patient management.

Purpose of the Study:

• To develop an advanced deep learning model for precise breast cancer mass classification in ultrasound images.
• To overcome the limitations of current imaging techniques in detecting subtle mass characteristics.

Main Methods:

• Proposed Dynamic Multiscale Fusion Transformer (DMFormer), a novel Transformer-based architecture.
• Integrated dynamic multiscale feature fusion with window and grid attention mechanisms.
• Enabled comprehensive capture of fine-grained tissue details and broader anatomical contexts.

Main Results:

• DMFormer achieved areas under the curve of 90.48% and 86.57% on two independent datasets.
• The model consistently outperformed state-of-the-art approaches, including CNNs, Transformers, and hybrid models.
• Demonstrated superior performance in differentiating benign from malignant breast masses.

Conclusions:

• DMFormer exhibits superior performance in ultrasound breast cancer detection via a dual-attention approach.
• The model effectively balances local and global feature processing with computational efficiency.
• Validated potential for enhancing accuracy and reliability in clinical breast cancer screening.