A multi-scale, multi-task fusion UNet model for accurate breast tumor segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is a leading cause of death in women globally.
• Accurate interpretation of small, complex breast tumors is challenging and time-consuming.
• Developing automated segmentation models is crucial for clinical assistance.

Purpose of the Study:

• To develop a multi-scale, multi-task deep learning model for breast tumor segmentation.
• To improve the accuracy and efficiency of breast tumor interpretation.
• To provide a tool to assist clinicians in breast cancer treatment.

Main Methods:

• Proposed MTF-UNet framework utilizing multi-scale, multi-level feature extraction with a single kernel size.
• Introduced an attention-based feature fusion block (ADF) for enhanced feature integration.
• Implemented a novel auxiliary task predicting pixel counts to aid segmentation.

Main Results:

• Achieved high Mean Intersection over Union (MIoU) scores: 90.45% (proprietary dataset), 89.84% (BUSI), and 92.84% (Kvasir-SEG).
• Ablation studies confirmed the effectiveness of individual components and the auxiliary prediction branch.
• Demonstrated superior performance compared to existing algorithms.

Conclusions:

• MTF-UNet shows effectiveness, adaptability, and robustness in medical image segmentation.
• The model has significant potential for advancement in breast tumor analysis.
• Code and data are publicly available for further research.