Research On Coronary Artery Image Segmentation Based On Limited Fine-labeled Data

Cardiovascular disease is a common disease that seriously threatens people.Coronary artery-caused diseases,especially,are an important cause of death all over the world.At the same time,due to the uneven distribution of professional doctors,many remote hospitals are facing difficulties in seeing doctors.How to combine medical and artificial intelligence to build an intelligent diagnosis system for cardiac coronary artery disease has become a subject of great research value,and the most important part is the segmentation of cardiac coronary image.The segmentation task requires the input of a coronary coronary image,and the network can output a pixel-level multi-label segmentation result,including different types of main blood vessels,unrelated blood vessels,guide wires,and backgrounds that can be seen in a specific posture.The scale of training data is significant in segmentation task,especially in segmenting the medical coronary artery angiograms.Traditional semantic segmentation networks have been restricted in this field,due to the specialty of cardiac coronary angiography data,that is,it is very difficult to balance the manual labeling costs and network accuracy.Based on this,a method for segmenting cardiac coronary images based on a small amount of fine-labeled data is proposed.The main works of this paper are as follows:firstly,the creation process and specifications of vascular image dataset in medical data are proposed in this paper,including the definition,generation steps and specifications of raw images,sketchy-labeled images,precisely-labeled images,and binary images of different postures;secondly,a new method to automatically generate so-called ’pseudo-precise’ label is proposed;finally,this paper optimizes the existing image segmentation network.Besides,based on the thought of "evolutionary network",the paper designs a complete training pipeline for cardiac coronary image segmentation network.All these methods can improve the performance of the networks on the premise of reducing labor costs as much as possible,and thus increase the F1-score by 4%-11%.