Physionet Challenge 2024
Training computer vision transformers to deskew images for ECG digitizaition
Project Overview
As part of the Physionet 2024 Challenge, I along with 4 other members in Edwards Lifesciences competed to develop a computer vision pipeline to convert paper ECGs into a digital format. My task was create a pipeline to deskew images of ECGs.
I used a computer vision transformer model, finetuning it based on synthetically generated scans of ECG images. The model was trained locally on a GPU, and then dockerized and deployed to the Physionet servers. Due to resource constraints, the model was not able to be used in the final product; however, the model did achieve 90% accuracy, and was explored as potential improvement to the rule-based method we used in our paper.
Publication
Our paper, "Fusion of Deep Learning and Rule-Based Techniques for Enhanced Paper-Based ECG Digitization" was accepted to the Computing in Cardiology 2024 Conference. The paper details the development of our pipeline, and was given the award of Best Preprint.
The paper was presented at the conference, and will be published in November.