DDUM: Deformable Dilated U-structure Module for coronary stenosis detection.

Area of Science:

• Cardiology
• Medical Imaging
• Artificial Intelligence

Background:

• Deep learning aids coronary artery disease diagnosis but struggles with limited, hard-to-label coronary angiography data, leading to low accuracy and poor generalization.
• Existing models often exhibit high false-positive rates and limited ability to adapt to new datasets.

Purpose of the Study:

• To introduce a novel Deformable Dilatable U-structure Module (DDUM) designed to specialize networks for coronary stenosis detection.
• To enhance the accuracy and generalization capabilities of deep learning models for analyzing coronary angiography.

Main Methods:

• Development of the Deformable Dilatable U-structure Module (DDUM).
• Integration and testing of DDUM with standard deep learning architectures (e.g., ResNet50 backbone with Faster R-CNN detector).
• Evaluation of performance improvements and generalization ability via transfer learning experiments.

Main Results:

• DDUM significantly improved model performance across various architectures.
• Application of DDUM to a ResNet50/Faster R-CNN model boosted mean average precision from 33.76% to 42.39% (a 25.56% increase).
• Transfer learning experiments confirmed DDUM's enhanced generalization capabilities.

Conclusions:

• The DDUM effectively improves coronary stenosis detection accuracy and model generalization.
• DDUM offers a specialized solution for deep learning in medical imaging, overcoming data limitations.
• Fine-tuning with DDUM reduces training costs and facilitates easier model deployment across devices.