MetaV: A Pioneer in feature Augmented Meta-Learning Based Vision Transformer for Medical Image Classification

Area of Science:

• Computer Vision
• Machine Learning
• Medical Imaging

Background:

• Image classification faces challenges with limited data, noise, and interpretability.
• Vision transformers offer potential but require substantial training data.
• Existing methods struggle with fine-grained classification and generalization.

Purpose of the Study:

• To introduce MetaV, an innovative meta-learning vision transformer for medical image classification.
• To enhance feature representation and mitigate overfitting in vision transformers.
• To address data scarcity and noise in medical datasets, especially for rare diseases.

Main Methods:

• Integration of meta-learning with a vision transformer architecture using N-way K-shot learning.
• Incorporation of deformational convolution and patch merging techniques.
• Application of augmentation methods like perturbation and Grid Mask.

Main Results:

• MetaV achieved high accuracies: 89.89% (Break His), 87.33% (ISIC 2019), 94.55% (SIPaKMed), and 80.22% (STARE).
• Demonstrated superior performance compared to conventional models.
• Established a new benchmark for meta-vision image classification.

Conclusions:

• MetaV effectively overcomes limitations of traditional vision transformers in medical image classification.
• The proposed model shows significant improvements in accuracy, generalization, and handling noisy, scarce data.
• MetaV offers a promising direction for advancing medical image analysis and diagnosis.