A Deep Semantic Segmentation Framework for Digitizing Paper Electrocardiograms with Data Augmentation using a Diffusion Model

Ji, Jiaqi

Etd

A Deep Semantic Segmentation Framework for Digitizing Paper Electrocardiograms with Data Augmentation using a Diffusion Model

Public Deposited

Electrocardiograms (ECGs) play a vital role in the diagnosis of cardiac diseases. ECGs have traditionally been analyzed manually by human experts, which is time consuming and yields inconsistent results. In electronic form, ECGs can be analyzed using modern methods such as machine learning. However, a large number of ECGs are currently stored in paper form, especially in developing countries. Moreover, paper ECGs are easily damaged, degenerate easily, and are difficult to preserve. In this thesis, the problem of digitizing ECGs to facilitate Machine Learning (ML) analyses is explored. The use of traditional image processing techniques for ECG digitization is challenging due to the wide variety of colors and signal variations in real ECGs. We propose a neural networks approach that uses the UNet deep semantic segmentation model to identify and segment regions of medical paper containing the ECG signal. To facilitate deep learning approaches, first we augmented the small ECG dataset by using a diffusion model to generate synthetic ECG signals. In rigorous evaluation, our best-performing model using the VGG16 model, achieved an F1 score of 92.52%, and mIoU of 86.88%, outperforming state of the art baseline models.

Creator