A Multi-Task Transformer With Local-Global Feature Interaction and Multiple Tumoral Region Guidance for Breast Cancer...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer remains a significant global health concern with high incidence and mortality.
• Accurate interpretation of breast ultrasound images is crucial for early diagnosis but demands extensive physician expertise.
• Current deep learning models often address segmentation and classification separately, missing key diagnostic information.

Purpose of the Study:

• To develop an advanced multi-task deep learning network for simultaneous breast tumor segmentation and classification.
• To effectively integrate local and global features from ultrasound images.
• To leverage specific tumor regions for improved diagnostic accuracy.

Main Methods:

• A dual-stream encoder (CNN and Transformer) was designed for hierarchical local-global feature interaction and fusion.
• A multi-tumoral region guidance module was implemented to capture non-local dependencies within tumor areas.
• The network was trained and evaluated on two breast ultrasound datasets.

Main Results:

• The proposed network demonstrated superior performance in both tumor segmentation and classification tasks compared to state-of-the-art methods.
• Achieved a significant improvement in diagnosis accuracy from 73.64% to 80.21% on an external validation dataset.
• Showcased strong generalization capability on unseen data.

Conclusions:

• The developed multi-task network effectively integrates local and global features for enhanced breast tumor analysis.
• The multi-tumoral region guidance module provides valuable interpretable cues for classification.
• This approach offers a promising advancement for automated breast ultrasound interpretation and diagnosis.