RTS-ViT: Real-Time Share Vision Transformer for Image Classification.

Area of Science:

• Computer Vision
• Medical Imaging
• Artificial Intelligence

Background:

• Vision transformers (ViTs) excel in image classification.
• Dual-branch ViTs improve feature generation through fusion.
• Retinal image classification requires robust feature representation.

Purpose of the Study:

• To develop an efficient dual-branch vision transformer for retinal image classification.
• To enhance feature learning and fusion mechanisms in ViTs.
• To improve classification accuracy while reducing computational complexity.

Main Methods:

• Proposed a dual-branch vision transformer with a Real-Time Share feature encoder.
• Processed image patches of base and large sizes through independent branches.
• Implemented multi-stage feature fusion using L-Times Attention Fusion (vector concatenation and element-wise addition).

Main Results:

• Achieved superior performance compared to Cross-ViT, with a 5.61% higher average TOP-1 Accuracy.
• Demonstrated lower FLOPs and fewer model parameters.
• Showcased strong self-learning capabilities without reliance on pre-trained weights.

Conclusions:

• The proposed dual-branch ViT with Real-Time Share feature significantly enhances retinal image classification.
• The L-Times Attention Fusion method offers efficient and effective feature integration.
• The approach presents a promising, resource-efficient solution for medical image analysis.