MEDUSA: Multi-Scale Encoder-Decoder Self-Attention Deep Neural Network Architecture for Medical Image Analysis.

Area of Science:

• Medical image analysis
• Artificial intelligence in healthcare
• Deep learning for diagnostics

Background:

• Subtle disease characteristics and overlapping appearances pose challenges in medical image analysis.
• Existing self-attention models often use multiple isolated attention mechanisms with limited capacity.
• A unified, high-capacity self-attention mechanism is needed for improved medical image interpretation.

Purpose of the Study:

• To introduce a novel multi-scale encoder-decoder self-attention (MEDUSA) mechanism for medical image analysis.
• To address the limitations of existing self-attention architectures in capturing complex disease patterns.
• To enhance the performance of deep learning models in identifying subtle and overlapping pathologies.

Main Methods:

• Development of MEDUSA, a unique "single body, multi-scale heads" self-attention mechanism.
• Implementation of a unified, high-capacity self-attention module integrated into an encoder-decoder framework.
• Evaluation of MEDUSA on challenging medical imaging datasets including COVIDx, RSNA RICORD, and RSNA Pneumonia Challenge.

Main Results:

• MEDUSA achieved state-of-the-art performance across multiple challenging medical image analysis benchmarks.
• The model demonstrated superior ability in discerning subtle disease characteristics and differentiating overlapping pathologies.
• Explicit global context and differing local attention contexts were enabled at multiple representational abstraction levels.

Conclusions:

• MEDUSA represents a significant advancement in self-attention for medical image analysis.
• The "single body, multi-scale heads" architecture effectively captures global and local contextual information.
• The publicly available MEDUSA model offers a powerful tool for improving diagnostic accuracy in medical imaging.