A naukri.com initiative
Arxiv
2d
58
Image Credit: Arxiv
Deep Learning-Based Digitization of Overlapping ECG Images with Open-Source Python Code
- Researchers have developed a deep learning-based method to digitize paper-based electrocardiogram (ECG) recordings, specifically addressing signal overlaps in single leads.
- A two-stage pipeline was proposed: first using a U-Net based segmentation network to isolate the primary ECG trace and then converting it into a time-series signal.
- The segmentation network achieved an IoU of 0.87 for fine-grained segmentation, outperforming existing methodologies.
- The digitization method showed superior performance to a baseline technique for both non-overlapping and overlapping ECG samples, with lower Mean Squared Error and higher Pearson Correlation Coefficient.
- For non-overlapping signals, the method achieved an MSE of 0.0010 and a rho of 0.9644 compared to the baseline's 0.0015 and 0.9366.
- On overlapping samples, the method achieved an MSE of 0.0029 and a rho of 0.9641, whereas the baseline had 0.0178 and 0.8676.
- This work presents an effective strategy to enhance digitization accuracy, particularly in the presence of signal overlaps in ECG recordings, enabling the conversion of analog ECG records into digital data for research and clinical use.
- The implementation code is available on GitHub at this repository: https://github.com/masoudrahimi39/ECG-code.
Read Full Article
3 Likes
For uninterrupted reading, download the app