TRAM-UNet: Transformer and Region Attention Module based U-Net for Breast Ultrasound Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of breast ultrasound images is vital for early breast cancer detection.
• Existing segmentation models face challenges in handling diverse lesion characteristics and refining boundaries.

Purpose of the Study:

• To introduce TRAM-UNet, a novel deep learning model integrating Transformer blocks and a Region Attention Module (RAM).
• To enhance the performance of breast ultrasound image segmentation using TRAM-UNet.

Main Methods:

• Development of the TRAM-UNet architecture, combining Transformer blocks with a Region Attention Module (RAM).
• Evaluation of TRAM-UNet on three distinct breast ultrasound datasets: BUS-BRA, BUSI, and BLUI.
• Comparative analysis against U-Net and U-Net + Transformer models.

Main Results:

• TRAM-UNet achieved superior average Dice scores: 88.56% (BUS-BRA), 84.68% (BUSI), and 83.96% (BLUI).
• The model demonstrated significant performance improvements over U-Net and U-Net + Transformer across all datasets.
• TRAM-UNet effectively refines boundaries and adapts to various lesion characteristics.

Conclusions:

• TRAM-UNet represents a significant advancement in deep learning for breast ultrasound image segmentation.
• The model's enhanced accuracy and adaptability show potential for improving automated breast cancer diagnosis.
• Further research can optimize TRAM-UNet for clinical application in breast cancer screening and diagnosis.