Real-time Segmentation of Coronary Artery Calcification Using Spatial Attention and Parallel Convolution

Semantic segmentation is an image recognition technique that classifies each pixel in an image. This technique is crucial in the fields of automated driving and medicine. Despite its significance, limited computational resources is available for its application. The current challenge involves to developing models that are not only accurate but also lightweight and fast. DDRNet (Deep Dual-Resolution Net) is a real-time segmentation method inspired by HRNet (High-Resolution Net), which is a highly accurate method with a multi-resolution substructure. DDRNet aims to achieve both accuracy and real-time performance reducing the number of resolutions used in the subnetworks and performing feature extraction more efficiently than HRNet. However, DDRNet is less accurate than non-lightweight models, especially in small object recognition. Therefore, this research aimed to improve the accuracy of DDRNet while maintaining its lightweight and high computation speed. We conducted experiments on a coronary calcification dataset using DDR-SAPC-U, which is a modified version of the proposed method specialized for medical imaging with the addition of a skip connection. Overall, DDR-SAPC-U demonstrated the highest accuracy compared to conventional medical methods and was relatively lightweight in terms of computational complexity. Compared to conventional methods, DDR-SAPC-U improved the accuracy of left anterior descending artery calcification by 7.59 points.