MedViT: A robust vision transformer for generalized medical image classification.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• Deep learning models like Convolutional Neural Networks (CNNs) are vital for automated medical diagnosis.
• Concerns exist regarding the vulnerability of these systems to adversarial attacks, potentially compromising patient safety.
• Existing models often struggle to balance computational efficiency with robustness against such threats.

Purpose of the Study:

• To develop a highly robust and efficient hybrid model combining CNNs and Transformers for medical image analysis.
• To enhance the reliability of automated medical diagnosis systems against adversarial attacks.
• To improve the generalization ability of diagnostic models on diverse medical datasets.

Main Methods:

• Proposed a CNN-Transformer hybrid architecture integrating CNN locality with Transformer global connectivity.
• Developed an efficient attention mechanism using convolution to mitigate the quadratic complexity of self-attention.
• Implemented a novel technique to learn smoother decision boundaries by augmenting shape information via feature permutation within mini-batches.

Main Results:

• The hybrid model demonstrated superior robustness and generalization capabilities compared to state-of-the-art methods.
• Achieved high performance on large-scale standardized MedMNIST-2D datasets.
• The proposed attention mechanism reduced computational complexity while maintaining effectiveness.

Conclusions:

• The developed CNN-Transformer hybrid model offers a robust and efficient solution for medical image diagnosis.
• The findings suggest a promising direction for creating more secure and reliable AI-driven medical systems.
• This approach effectively addresses the vulnerability of deep learning models to adversarial attacks in critical healthcare applications.