Self-supervised learning for breast cancer detection: A review.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Machine learning for diagnostics

Background:

• Breast cancer remains a leading global health concern, necessitating advancements in early detection and diagnosis.
• Deep learning (DL) shows potential in computer-aided detection (CAD) but requires extensive labeled datasets.
• Self-supervised learning (SSL) offers a solution by utilizing unlabeled data for robust feature learning.

Purpose of the Study:

• To review the application of SSL in breast cancer detection across screening, diagnosis, grading, and staging.
• To analyze SSL's impact on various imaging modalities like mammography, ultrasound, MRI, and histopathology.
• To identify gaps and future directions for SSL in breast cancer imaging.

Main Methods:

• Comprehensive literature review of SSL applications in breast cancer imaging.
• Analysis of SSL's role in reducing annotation burden and improving model generalization.
• Focus on modalities including mammography, digital breast tomosynthesis, ultrasound, MRI, and histopathology.

Main Results:

• SSL shows success in mammography and ultrasound for early detection and in MRI/histopathology for lesion characterization.
• SSL effectively reduces annotation demands and enhances generalization across different data domains.
• A significant gap exists in SSL research for PET imaging within breast cancer diagnostics.

Conclusions:

• SSL presents a transformative approach to breast cancer imaging, offering scalable solutions with reduced reliance on expert annotations.
• Future research should explore SSL in underexplored modalities like PET and focus on multi-modal fusion and explainability.
• SSL holds promise for improving breast cancer detection, characterization, and prognostication.