Research On Coronary Angiography Vessel Segmentation Algorithm Based On Knowledge Distillation

In recent years,the problem of increased incidence of cardiovascular diseases such as coronary heart disease and acute coronary syndrome has become more serious.The automatic and accurate segmentation of cardiac coronary angiography blood vessels by computer deep learning algorithm is an effective auxiliary means for the prevention and diagnosis of cardiovascular diseases.However,high-precision deep learning-based segmentation algorithms often need to be trained on large Graphics Computing Servers(GPUs),which is not conducive to deployment and implementation in actual medical scenarios.In response to the above issues,this paper proposes a novel knowledge distillation algorithm,which can significantly improve the performance on the task of vessel segmentation without increasing the computational cost of the algorithm model.This research has theoretical basis and the great practical value.To be specific,the main contributions could be listed as follows:1.In this paper,a pixel-level fine-grained cardiac coronary vessel segmentation dataset is constructed and annotated,and the dataset is screened and augmented to facilitate the experimental demonstration of the proposed algorithm model.2.This paper proposes a novel semantic feature-level similarity knowledge distillation module.This module can encode and reconstruct the features of both the teacher network and the student network from the feature latent space dimension.It helps to improve the extraction ability and robustness of the lightweight student network model for semantic information.3.A novel semantic pixel-level similarity knowledge distillation module is proposed which could utilize the last semantic segmentation output of the teacher network to refine the prediction results of the student network from the pixel-level dimension in a fine-grained manner.4.In this paper,various generalization verification experiments are carried out on the optimized lightweight model.It is proved that the knowledge distillation algorithm proposed in this paper has achieved excellent performance on various lightweight models.This also reflects the high practical value of the algorithm.The proposed knowledge distillation algorithm can be applied to a variety of lightweight networks without increasing the computational cost of the model.Extensive ablation and generalization experiments are performed on a self-constructed cardiac coronary angiography vessel segmentation dataset to demonstrate that the knowledge distillation algorithm proposed in this paper is superior to existing classical and stateof-the-art knowledge distillation algorithms.