Arrhythmia detection model using modified DenseNet for comprehensible Grad-CAM visualization. (March 2022)

• Record Type: Journal Article
• Authors: Kim, Jin-Kook; Jung, Sunghoon; Park, Jinwon; Han, Sung Won
• Is Part Of: Biomedical signal processing and control. Volume 73(2022)
• Publication Date: 2022-03
• Subjects: ECG Electrocardiogram -- DenseNet Densely connected convolution-nal network -- Grad-CAM Gradient-weighted class activation mapping -- CNN Convolutional neural network -- CAM Class activation mapping -- GAP Global average pooling -- AF Atrial flutter; Arrhythmia classification -- Electrocardiogram -- Class activation mapping -- Convolutional neural network; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103408
• Languages: English
• ISSNs: 1746-8094

Abstract

Highlights: Deep Learning was used to generate a model that detects arrhythmia using ECG. The basis for judgement was difficult to understand in the basic model structure. This study improve the visualization of Grad-CAM without compromising classification accuracy. This study allows us to visualize irregular intervals or shapes of electrocardiogram. An interpretable model will enable doctors to gain trust in medical deep learning. Abstract: Diagnosing arrhythmia is difficult, requires significant efforts. Because arrhythmia can be associated with serious diseases, it is important to classify arrhythmia patients with high accuracy, and the basis for the classification model's judgment should be properly demonstrated. Traditional algorithm methods are less accurate, and simply using a high-accuracy image classification deep learning model yields incomprehensible results when the model is visualized with gradient-weighted class activation mapping (Grad-CAM). We want to achieve high-performance deep learning models can also comprehensible visualization. To obtain this, two hypotheses about Grad-CAM were established and the experiment was conducted. As a result, a method that could clearly visualize the response area using Grad-CAM with a higher classification performance of 0.98 accuracy is created.