Complex edge extraction and plaque recognition of intravascular ultrasound images based on multi-task learning

Intravascular ultrasound (IVUS) can accurately reflect the nature, severity and pathological composition of cardiovascular diseases. It is considered as a new "Gold standard" for the detection of cardiovascular diseases. IVUS image edge extraction and plaque recognition are the basis of the Quantitative analysis of atherosclerotic plaque, and are also valuable for the qualitative study of subsequent vascular elastic analysis and three-dimensional reconstruction. most of the current methods of edge extraction and plaque recognition in IVUS images are algorithmically complex or focus is always on a single task, thus neglecting other favorable optimization metrics. The multi-task learning method proposed in this paper can automatically extract the edge of lumen-Intima (LI) and media-Adventitia (MA), at the same time, it can identify the type of plaque, which is convenient and efficient. In addition, we propose a task interaction module to fuse the features of the two branches of the network to maximize the sharing of information between different tasks and extract their own useful features. An improved efficient channel attention (ECA) module is added to enhance the effective feature extraction. The experimental results show that this method not only has good learning effect, but also is more robust.