Challenge : Extract the ECG time-series data from scans and photographs of paper printouts of the ECGs.
Types of training images classes:
Original color ECG image generated by ECG-image-kit.
Image printed in color and scanned in color.
Image printed in color and scanned in black and white.
Mobile photos of color printed images.
Mobile photos of ECGs on the screen of laptop.
Mobile photos of stained and soaked printed ECGs.
Mobile photos of printed ECGs with extensive damage.
Scans of printed ECG images with mold in color.
Scans of printed ECG images with mold in black and white
A four stage pipeline:
-
Classification: We developed an image classifier (ResNet-18, fine-tuned on our own categorized ECG images) predicted what type of scan the image was according to the categories and routed it to the right preprocessing branch.
-
Preprocessing: We had different preprocessing pipelines for each class of image with necessary screen content cropping, glare reduction, perspective distortion correction using contour detection, contrast enhancement via CLAHE and unsharp masking, and red gridline suppression.
-
Lead detection: We trained a YOLOv7 model which detected bounding boxes for each of the 12 lead regions on the preprocessed image.
-
Lead extraction: For each cropped lead region, we removed the ECG grid using morphological operations, then extracted the waveform using a proximity-based column walker that tracked the signal pixel-by-pixel and interpolated gaps.