Robust Visual Transformers for Medical Image Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging Analysis

Background:

• The Vision Transformer (ViT) architecture offers scalability and global attention, driving interest in computer vision.
• ViT's adaptability makes it suitable for various applications, but its performance on complex, data-scarce medical imaging tasks needs enhancement.

Purpose of the Study:

• To investigate the performance boundaries of the Vision Transformer architecture.
• To develop novel techniques for improving ViT performance on complex tasks, specifically medical imaging datasets.
• To address challenges like high variability, class imbalance, and limited sample sizes in medical imaging.

Main Methods:

• Proposed a set of mixed regularization and augmentation techniques tailored for complex computer vision tasks.
• Introduced a novel loss function to improve model training stability and performance.
• Developed a smoothly differentiable activation function to enhance training dynamics.

Main Results:

• Incorporation of the proposed techniques significantly improved model performance on medical imaging datasets.
• Observed enhanced training convergence, indicating greater stability and efficiency.
• Demonstrated the effectiveness of mixed regularization and augmentation for challenging datasets.

Conclusions:

• The developed techniques effectively enhance Vision Transformer performance in complex scenarios, particularly medical imaging.
• The novel loss function and activation function contribute to more stable and efficient model training.
• This work provides a valuable framework for applying ViTs to high-variability, limited-sample medical imaging tasks.