A deep learning-based approach for scanned electrocardiogram image digitization using generative models
Loading...
Author(s)
Kodthalu Shivashankara, Kshama
Advisor(s)
Sameni, Reza
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
The use of deep learning in the diagnosis of cardiovascular diseases has become a prevalent research topic in recent times. The electrocardiogram (ECG) is a medical record frequently used in developing deep-learning algorithms to predict the presence of cardiovascular diseases. However, many ECGs are needed to train deep learning models, and many such records exist only in paper formats, making them unsuitable for training. We propose a solution to generate millions of synthetic scanned paper ECG records and digitize paper ECGs to digital data fully automatically using deep learning models. To account for the variability in paper ECG images, we present a deep learning-based digitization solution. The proposed method employs a Denoising Convolutional Neural Network (DnCNN) to perform grid removal and image denoising of the paper ECG image. Evaluation of the algorithm has been done on 1000 synthetically generated images by computing the Signal-to-Noise Ratio (SNR) of the digitized time-series data with respect to the ground-truth time-series data, achieving an average SNR of 27.03 dB per image. For a clinical evaluation, the error between the estimated QT-intervals of the digitized and ground-truth time-series data is measured. The results demonstrate the effectiveness of a deep learning-based pipeline in achieving accurate digitization of paper ECGs while also highlighting a generative approach to digitization.
Sponsor
Date
2023-07-25
Extent
Resource Type
Text
Resource Subtype
Thesis