A Semantic Segmentation-Based Digitization of ECG Papers

The electrocardiogram (ECG) is crucial to identify biomarkers of cardiovascular diseases (CVD) in a non-invasive and non-expensive manner. While digital ECGs have become more prevalent and offer many advantages, the use of ECGs papers may still be necessary in certain situations. Healthcare providers must weigh the benefits and limitations of each method and choose the most appropriate option. Once ECG papers are converted into digital format, they can be used to teach machine learning models how to accurately diagnose heart conditions. By using a collection of digital ECGs alongside their corresponding diagnoses, healthcare providers can help train these models to automate the diagnostic process, making it faster and more efficient. This paper proposes a digitization pipeline that uses a combination of machine learning and image processing techniques. The system identifies the regions of interest on the scanned papers and segments the ECG signal from the background, by using a Random Forest model trained using a Gabor filter bank. Finally, the ECG waveform in millivolt units is extracted through an iterative process that uses the ECG baseline as a reference. After testing the proposed digitization pipeline on two sets of images, the results were accurate (MSE = 0.021). Also, the estimations were closely matched the actual ECG data (Correlation Coefficient = 0.861).