Symmetry-based regularization in deep breast cancer screening.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Breast cancer is a leading cause of mortality in women, necessitating improved screening methods.
• Neural network-based computer-aided diagnosis (CAD) systems show promise for early breast cancer detection.
• Data scarcity and annotation challenges limit the effectiveness of current AI models in medical imaging.

Purpose of the Study:

• To develop and unify regularization methods that leverage domain-specific symmetries for breast cancer screening AI.
• To enhance the generalization capabilities of neural networks in the face of limited medical data.
• To provide a framework for more data-efficient AI development in medical imaging.

Main Methods:

• A unified approach to regularization incorporating known symmetries into neural network models.
• Implementation of three strategies: data augmentation, loss function invariance promotion, and equivariant architectures.
• Empirical validation across diverse datasets, scenarios, and model architectures.

Main Results:

• Symmetry-based regularization significantly improves model generalization to unseen breast cancer images.
• The proposed methods demonstrate effectiveness across various datasets and neural network architectures.
• The techniques are readily applicable to most AI modeling settings.

Conclusions:

• Symmetry-based regularization is a powerful strategy to overcome data scarcity in medical AI.
• These methods enhance the reliability and efficiency of AI for breast cancer screening.
• The presented principles and techniques can advance data-efficient AI in broader medical imaging applications.