Transformer attention fusion for fine grained medical image classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Diabetic retinopathy (DR) is a leading cause of preventable blindness, necessitating precise and timely diagnosis.
• Automated DR classification faces challenges like irregular lesions, class imbalance, and variable image quality, impacting accuracy.
• Fine-grained visual classification is crucial for detecting subtle abnormalities in medical images.

Purpose of the Study:

• To develop an advanced AI model for automated, accurate, and early detection of diabetic retinopathy.
• To address the limitations of existing DR classification systems, including data imbalance and image quality inconsistencies.

Main Methods:

• Introduced MSCAS-Net (Multi-Scale Cross and Self-Attention Network) utilizing a Swin Transformer backbone.
• Employed multi-scale feature extraction (12x12, 24x24, 48x48) to capture both local and global image details.
• Integrated self-attention for intra-scale spatial connections and cross-attention for inter-scale feature matching.

Main Results:

• MSCAS-Net achieved high accuracy on benchmark datasets: 93.8% (APTOS), 89.80% (DDR), and 86.70% (IDRID).
• The model effectively handles imbalanced datasets and inconsistent image quality without data augmentation, learning stable features.
• Demonstrated superior performance in detecting subtle and significant lesions indicative of DR.

Conclusions:

• MSCAS-Net represents a breakthrough in automated DR diagnostics, offering high precision and interpretability.
• The AI model functions as an efficient clinical decision support system for early DR detection and management.
• Fine-grained visual classification methods, as implemented in MSCAS-Net, significantly benefit early-stage DR detection and treatment.