Semi-supervised vision transformer with adaptive token sampling for breast cancer classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Machine learning (ML) and deep learning, particularly convolutional neural networks (CNNs), have advanced computer-aided diagnosis (CAD) for breast cancer (BC) detection and classification.
• Supervised learning approaches have dominated ML-based BC detection, achieving expert-level accuracy.

Purpose of the Study:

• To propose a novel semi-supervised learning framework for breast cancer detection utilizing the Vision Transformer (ViT) model.
• To enhance the robustness and performance of BC detection models by integrating supervised and consistency training.
• To explore the efficacy of ViT, a model less commonly applied to BC detection, in improving diagnostic accuracy.

Main Methods:

• Development of a custom semi-supervised learning framework based on the Vision Transformer (ViT).
• Implementation of a unified training procedure combining supervised learning and consistency training for enhanced model robustness.
• Integration of an adaptive token sampling technique to focus on the most significant image features.

Main Results:

• The proposed ViT-based semi-supervised framework demonstrated consistent performance improvements over CNN baselines.
• Validation on both ultrasound and histopathology image datasets confirmed the method's effectiveness in breast cancer detection.
• The adaptive token sampling technique contributed to significant performance gains.

Conclusions:

• The Vision Transformer (ViT) shows significant promise for breast cancer detection, offering superior performance compared to traditional CNN models.
• Semi-supervised learning, combined with adaptive token sampling, provides a robust and effective approach for improving BC diagnostic accuracy.
• The developed framework offers a valuable advancement in AI-driven breast cancer diagnosis.