Driving ECG Digitization - Using Techniques from Autonomous Driving to Detect Regions of Interest in ECG Images

Digitizing ECG images, i.e. reconstructing signal waveforms from paper printouts or screen captures could make AI-based diagnosis a reality, especially in the global south.We, team RoadRuNNers, transfer algorithms borrowed from autonomous driving -originally designed for road lane detection -to detect the region of interest (ROI) in ECG images and take them as input for a convolutional neural network (CNN).First, we invert the grayscale of the ECG image in order to make the ECG lines appear white before a dark background, similar to a road lane.Subsequently, we adopt open source scripts that iteratively detect white lines and their orientation by randomly choosing thresholds for binarization, Gaussian blur kernel size, and Canny edge detection.The detected 4 -6 ROIs are assumed to show ECG signals when they are in parallel and equidistant to each other and span at least half of the image size.The ROIs then serve as input for a CNN that processes the 4 binary images to reconstruct 4 10s of ECG signal.Our approach in 'Digitization and Classification of ECG Images: The George B. Moody PhysioNet Challenge 2024' achieved a digitization score of -0.071 (rank 10/16) on the hidden test data.Despite improvable performance levels, we demonstrated the potential of adapting techniques from autonomous driving for ROI detection.