结合迁移学习与深度卷积网络的心电分类研究

为解决一维深度卷积网络（1D-DCNN）在心电分类方面存在的多类疾病识别不准、难以提取最佳特征等问题，提出一种结合迁移学习与二维深度卷积网络（2D-DCNN）直接识别心电图像的方法。首先，截取R波前后75 ms内的心电信号，并将一维心电电压信号转化为二维灰度图像信号。接着，构建2D-DCNN对心电节拍样本进行分类训练，权值初始化采用在ImageNet大规模图像数据集上进行预训练的AlexNet参数值。本文提出方法在MIT-BIH心电数据库上进行性能验证，其准确率达到98%，并在不同信噪比下保持较高的准确率，证明了所述模型在心电分类上具有良好的鲁棒性。为了验证2D-DCNN的识别性能，实验部分与采用不同激活函数的1D-DCNN、近些年性能较好的深度学习方法进行比较。量化结果表明，结合迁移学习和2D-DCNN方法，比最优1D-DCNN算法，其准确率提升2%、敏感度提升0.6%、特异性提高4%；在二分类与多分类任务中，均好于现有的其他算法。

ECG classification based on transfer learning and deep convolution neural network

Abstract: One-dimensional deep convolution neural networks (1D-DCNN) for electrocardiogram (ECG) classification shows limitations in identifying various diseases and extracting the best morphological features. Herein, a method combining transfer learning and two-dimensional deep convolution neural network (2D-DCNN), AlexNet, is proposed to identify ECG images directly. Firstly, the ECG signals within the 75 ms before and after R wave were intercepted, and one-dimensional ECG voltage signals were converted into two-dimensional grayscale image signals. Then, a 2D-DCNN based on AlexNet was established to classify the ECG heartbeat samples. The weights were initialized by parameters which were pre-trained on Alexnet using a large-scale dataset ImageNet. The proposed method achieved an accuracy of 98% on the MIT-BIH arrhythmia database, and maintained a high accuracy at different signal-to-noise ratios, which verified the good robustness of the proposed method in ECG classification. The proposed method was also compared with 1D-DCNN using different activation functions and other deep learning methods with favorable performances to evaluate the performance of 2D-DCNN. The quantitative results demonstrated that compared with the optimal 1D-DCNN, the proposed method combining transfer learning with 2D-DCNN improves the accuracy rate, sensitivity and specificity by 2%, 0.6% and 4%, respectively, and that the proposed algorithm is better than other existing algorithms in both binary/multi-class classification tasks.