Research On ECG Classification Network Model Pruning Algorithm Based On Score-CAM Framework

Objective: At present,a large number of ECG automatic diagnosis algorithms based on deep learning have been able to achieve better classification results on public data sets,but they are difficult to be applied in practice due to the large number of model parameters and high complexity.The pruning algorithm can greatly reduce the number of model parameters and reduce the complexity of the model without reducing the accuracy of the model.However,in the pruning process of the current pruning algorithm,some key features learned by the original model for classification decision will be pruned,which makes the visualization results of the model before and after pruning different,and also makes the simplified model obtained by pruning operation unreliable.The purpose of this paper is to design a pruning algorithm applied to the ECG classification network model,reduce the difference between the visualization results of the network model before and after pruning,and make the simplified model after pruning more reliable.Methods: This study proposes a pruning algorithm for ECG classification network model based on the Score-CAM framework.Using the visualization results of the original network model as a reference,calculate the similarity between the activation map of the convolution kernel class and the visualization results of the original network model,trim the convolution kernel corresponding to the lower similarity characteristic graph according to the pruning ratio,and fine-tune the network to restore the model accuracy,and obtain the final simplified model.Results: In this paper,weighted L2 norm pruning,random pruning,Taylor pruning,Apoz pruning methods are used as comparison algorithms.Experiments show that when the pruning ratio is 70%,the accuracy of the proposed method in the MIT-BIH dataset reaches 99.09%,and the F1 score reaches 99.08%,which is slightly higher than the 98.76% accuracy of the original VGG11 network and 98.75% F1 score.Compared with other pruning algorithms,under the condition that the generalization performance of the model is not reduced,the difference between the visualization results of the model before and after pruning is smaller.Conclusion: This paper is based on the Score-CAM framework,taking the visualization results of the pre-training model as a reference,and combining with the convolution kernel class activation map to measure the importance of the convolution kernel and complete the pruning.Compared with other pruning algorithms,the method proposed in this paper can make the difference between the visualization results of the model before and after pruning smaller without reducing the generalization performance of the model after pruning.