Aim: While traditional paper-based electrocardiograms (ECGs) continue to be prevalent in clinical settings, their digital conversion and subsequent data analysis are essential for the effective screening of cardiovascular diseases. We propose an improved deep learning algorithm and signal reconstruction method based on the Wave-GAN and ResNet-50 frameworks, achieving two sub-tasks of high-fidelity reconstruction of electrocardiogram waveforms and automatic classification.
Methods: This research innovatively introduces the Wave GAN framework based on conditional generative adversarial networks, which are synergized with the Hough transform to correct image rotation, thus producing high-fidelity ECG waveforms through an adversarial optimization of generator and discriminator networks. Furthermore, for waveform digitalization, we employ image processing techniques involving histogram analysis and contour detection to delineate regions of interest (ROI) precisely. By applying a scaling factor, the pixel coordinates in the binarized ROI area are converted into actual physical dimensions for accurate waveform digitalization. In terms of classification, we utilize Haar wavelet downsampling alongside several feature engineering strategies to suitably format ECG images for processing through an improved ResNet-50 model.
Results: Our team TimeBeater's methodology yields a signal-to-noise ratio of -18.123 in the reconstruction of waveforms. It attains a F-measure of 0.421 on the hidden validation set, signifying substantial improvements over traditional techniques.
Conclusions: This study enhances the effectiveness of electrocardiogram analysis by combining Wave-GAN and Hough transform for accurate waveform digitalization and using an optimized ResNet-50 model for superior classification, thereby offering a more reliable and efficient diagnostic approach.