[Coronary artery segmentation based on Transformer and convolutional neural networks dual parallel branch encoder...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Research

Context:

• Manual segmentation of coronary arteries in computed tomography angiography (CTA) is time-consuming and prone to errors.
• Existing deep learning models struggle with accurate coronary artery segmentation due to complex anatomical variations.

Purpose:

• To introduce a novel deep learning segmentation model, DUNETR (double parallel encoder u-net with transformers), for improved 3D coronary artery segmentation.
• To leverage a dual-encoder architecture combining Transformers and Convolutional Neural Networks (CNNs) for comprehensive feature extraction.

Summary:

• DUNETR utilizes a dual-encoder approach: a Transformer encoder for global, multi-scale features and a CNN encoder for local features.
• A noise reduction feature fusion (NRFF) module integrates these complementary features before passing them to the decoder.
• Experimental results show DUNETR achieved a Dice similarity coefficient of 81.19% and a recall of 80.18%, outperforming existing methods.

Impact:

• DUNETR significantly enhances the accuracy and efficiency of 3D coronary artery segmentation, surpassing conventional deep learning techniques.
• The model's architecture offers a promising new approach for segmenting other complex vascular structures in medical imaging.
• Improved segmentation accuracy can lead to better diagnosis and treatment planning for cardiovascular diseases.