Is Attention all You Need in Medical Image Analysis? A Review.

Area of Science:

• Artificial Intelligence in Medicine
• Deep Learning for Medical Imaging

Background:

• Medical imaging data constitutes nearly 90% of healthcare information, crucial for diagnosis and treatment.
• Convolutional Neural Networks (CNNs) excel at local feature extraction in medical image analysis (MIA) but struggle with global context.
• Transformers offer global context modeling but require extensive data and computational resources.

Approach:

• This systematic review surveys hybrid CNN-Transf/Attention models in MIA.
• It analyzes key architectural designs, breakthroughs, and challenges.
• An analysis framework for generalization is proposed to stimulate data-driven domain adaptation methods.

Key Points:

• Hybrid models synergize CNNs' local feature learning with Transformers/Attention's global context modeling.
• These architectures aim to overcome limitations of pure CNNs and full Transformers in MIA.
• The review highlights the growing trend and impact of hybrid models across various MIA applications.

Conclusions:

• Hybrid CNN-Transf/Attention models represent a significant advancement in medical image analysis.
• Further research is needed to fully exploit their potential for generalization and clinical impact.
• The proposed framework can guide the development of novel domain generalization and adaptation techniques.