ecg-image-generator

ecg-image-gen

Generating realistic ECG images from time-series data

This folder contains tools for generating realistic ECG images from time-series data, creating synthetic ECGs on standard paper-like backgrounds with genuine printing and scanning artifacts. Our approach adds distortions such as handwritten notes, wrinkles, creases, and perspective transforms. These images are ideal for producing large sets of ECG images for the development and evaluation of machine and deep learning models in ECG analysis.

The process of scanning and digitizing ECG images is governed by some fundamental limitations and requiements rooting in signal and image processing theory. A short overview of these concepts is available in a brief document found here.

Release History

• January 2024, initial public release

Installation

• Setup with Conda env:

  conda env create -f environment_droplet.yml
  conda activate myenv
  

• Set up with pip:

  1. Install all the dependencies:

     pip install -r requirements.txt
     

  2. If you will be using the handwritten text distortions feature, install sciSpacy with the following command:

     pip install https://s3-us-west-2.amazonaws.com/ai2-s2-scispacy/releases/v0.4.0/en_core_sci_sm-0.4.0.tar.gz
     

     Note that the requirements file has been compiled for python versions >= 3.8.11 and < 3.11

Running the pipeline

• The python script to generate the ECG images requires two mandatory parameters: the path to the input directory with the ECG time-series data and its header file (in PhysioNet's WFDB format) and the path to the output directory to store the synthetic ECG images. Here is an example case of running the python script with only the mandatory arguments:

  python gen_ecg_images_from_data_batch.py -i <path_to_input_directory> -o <path_to_output_directory> --print_header

• The gen_ecg_images_from_data_batch.py script produces the following outputs in each iteration:

  • Synthetic ECG image: Includes ECG signals from all leads, gridlines, and the name of each ECG lead.
  • Text and lead bounding box (optional): A CSV file detailing the grid size (xgrid and ygrid) for every generated image. This information can be used as ground truth for training machine learning and deep learning models.

• Below are sample synthetic ECG images generated from sample records of the PhysioNet PTB-XL dataset.

  Note: The ECG images generated for the format 3 by 4 (4 columns), the ECG signals for every column have been sampled in a shifted fashion, i.e. the leads in the first column (I, II, III) is from the first 2.5 seconds, the second column (aVR, aVL, aVF) is from the second 2.5 seconds, the third column (V1, V2, V3) is from the third 2.5 seconds segment and the fourth column (V4, V5, V6) is from the fourth 2.5 seconds segment of the respective lead of the ECG signal.

Generating distortionless ECG

The basic mode of the tool creates ECG images without distortions. The mode of operation and generated outputs can be configured using these command-line flags:

• -se: Seed controlling all the random parameters; type: int

• -r: Resolution with default being 200; type: int

• --pad_inches: Padding of white border along the image with default padding of 0 inches; type: int

• --print_header: Add text from header file on all the generated images; default: False

• --num_columns : Number of columns of the ECG leads. The default(-1) will plot a single column for 2 lead data and 4 columns for the 12 or any other number of lead data. Default: -1; type: int

• --full_mode: Sets the lead to add at the bottom of the paper ECG as a long strip obtained from the WFDB record's .hea header file, if the lead II is not available plots the first lead from the header file; default: 'II'; type: str

• --num_images: Number of ECG images to be generated; default: all files in the input directory; type: int

• --deterministic_lead: Remove lead names from all generated images, default=False.

• --random_resolution: Generate random resolutions of images, if True resolution is randomly picked from the range [50, -r] else every image is generated at the -r resolution; default: False

• --random_padding: Generate random padding widths on images, if True pad inches is randomly picked from the range [0, --pad_inches], else every image is padded with --pad_inches; default: False

• --random_dc: Add ECG calibration pulse to a random number of generated images. The parameter is the probability of the images having the calibration pulse; type: Float, default: 0 (no calibration pulse). Set to 1 to add the pulse to all images. Make sure that --random_dc $\in$ [0, 1].

• --random_grid_present: Probability of the generated images having the ECG paper grid; type: Float, default: 1 (adds the grid to all images). Make sure that --random_grid_present $\in$ [0, 1]. When 0, the images do not have the background grid.

• --random_add_header: Probability of adding printed text to a random set of images; type: Float, default: 0 (no text added). Make sure that --random_add_header $\in$ [0, 1]. If --print_header is True, code prints text on all the images regardless of the --random_add_header attribute.

• --random_bw: Make random set of images black and white controlled by this parameter; type: Float, default: 0 (generates colored ECG). Make sure that --random_bw $\in$ [0, 1].

• --standard_grid_color: Color of the grid lines, 1: brown, 2: pink, 3: blue, 4: green, 5(Default): red .Make sure that standard_grid_color $\in$ [1, 5], type: int

• --random_grid_color: Generates random colors for the gridlines, If '--random_bw > 0, then the color of gridlines for the non black and white ECG images is picked randomly. Default: False.

• --store_text_bounding_box: Store bounding box coordinates for the lead names in a text file in the folder output_directory/text_bouding_box; default: False.

• --bbox: Store bounding box coordinates for every individual ECG lead signal in a text file in the folder output_directory/lead_bouding_box; default: False.

• store_config: Store config information for each image in a json file, Refer to template.json for the template json file. Default: False. The json file has following attributes:

  • x_grid: Number of pixels per 200ms of the grid on the image.
  • y_grid: Number of pixels per 0.5mV of the grid on the image.
  • text_bounding_box_file: Path to the bounding boxes for the lead names if the store_text_bounding_box is True else ''
  • lead_bounding_box_file: Path to the bounding boxes for the ECG leads if the bbox is True else ''.

  Example:

  python gen_ecg_images_from_data_batch.py -i <path_to_input_directory> -o <path_to_output_directory> -se 10 --store_text_bounding_box --bbox --random_add_header 0.8 --random_dc 0.5

Adding distortions to the synethic images

• Text distortions

  Scanned ECG images often contain handwritten notes by physicians, sometimes overlapping the ECG traces. Our toolkit simulates this by using a dictionary of relevant keywords, which are randomly placed on the ECG images. We gathered medical texts related to ECG and cardiovascular diseases and employed Natural Language Processing (NLP) models to extract biomedical phrases and keywords. These were transformed into handwritten-style images using pretrained models and overlaid on the ECG images.

  For NLP, we utilized the Python-based en_core_sci_md model from sciSpacy for tokenization, parts of speech tagging, dependency parsing, and named entity recognition. The SpaCy model was retrained with our medical texts, focusing on ECG and cardiocascular context. We also retrained the dependency parser and parts of speech tagger using the McClosky and Charniak treebank, based on the GENIA 1.0 corpus.

  Our toolbox generates synthetic ECG images by parsing words from text files or online sources using the BeautifulSoup library, tagging them, and identifying ECG-related keywords with named entity recognition. These keywords are then converted into handwritten text using a pretrained Recurrent Neural Network (RNN) transducer-based model with a soft window.

  Adding the --hw_text flag to the python command provides this feature. Furthermore, following attributes specific to the text can be adjusted:

  • -l: URL to capture relevant ECG-related text for generating handwritten text artifacts; default: https://www.physionet.org/content/ptbdb/; type: str
  • -n: Number of handwritten words to add; default: 5; type: int
  • --x_offset: Defines the horizontal offset (in pixels) of the placed words from the image's border; default value: 30; data type: int.
  • --y_offset: Sets the vertical offset (in pixels) of word placement from the image's top border; default value: 30; data type: int.
  • --deterministic_offset: Use the provided offset parameters deterministically(Text is printed at [x_offset, y_offset]). If not, randomizes the text position based on x_offset and y_offset (x coordinate range:[1, x_offset+1], y coordinate range: [1, y_offset+1]); default: False
  • --deterministic_num_words: Uses the provided number of words deterministically. If False, it takes the number of words as a range and adds random number of words; default: False
  • --deterministic_hw_size: Uses a fixed handwriting size for the handwritten text artifacts added; default: False

  Example:

  python gen_ecg_images_from_data_batch.py -i my_input_dir -o my_output_dir --hw_text -n 4 --x_offset 30 --y_offset 20 -se 10 --random_grid_color

  • Below are sample synthetic ECG images with handwritten text generated from sample records of the PhysioNet PTB-XL dataset using the script above.

  Adding text artifacts is a computationally expensive process and should be run with GPU machines for large scale dataset generation.

• Wrinkles and creases

  Creases are simulated using Gaussian-blurred lines, evenly spaced to mimic paper fold creases. Gaussian blurring, a common image augmentation technique for smoothing effects, is applied to these lines. This blurring enhances realism by creating a shadow effect in the creases, common in scanned images or real paper ECG. For wrinkles, which are essentially textures, advanced texture synthesis methods like image quilting are used.

  Add --wrinkes to the python command to add wrinkle distoritions to the images. Furthermore following attributes specific to the wrinkles can be adjusted:

  • -ca: Crease angle (in degrees) with respect to the image; default: 90
  • -nv: Number of creases to add vertically; Default: 10
  • -nh: Number of creases to add horizontally; default: 10
  • --deterministic_angle: Chooses a fixed crease angle for all images; default: False
  • --deterministic_vertical: Adds the given number of vertical creases deterministically to all images; default: False
  • --deterministic_horizontal: Adds the given number of horizontal creases detereministically to all images; default: False

  Example:

  python gen_ecg_images_from_data_batch.py -i my_input_dir -o my_output_dir --wrinkles -ca 45 -se 10 --random_grid_color

  • Wrinkle and creases distortion on synthetic images generated from the PhysioNet PTB-XL

• Augmentation and noise

  Add --augment to the python command to add augmentations to the images. Furthermore following attributes specific to the wrinkles can be adjusted:

  • -rot: Rotation angle by which images can be rotated; default: 0; type: int
  • -noise: Noise levels to be added; default: 50; type: int
  • -c: Percentage by which image will be cropped; default: 0.01; type: int
  • -t: Colour temperature changes to be added to the image; default: 40000; type: int
  • --deterministic_rot: Adds the given amount of rotation to all images deterministically. If False, chooses rotation angles randomly in the given range; default: False
  • --deterministic_noise: Adds the noise level given detreministcally to all images. If False, adds random amounts of noise in the given range; default: False
  • --deterministic_crop: Adds the given level of crop to all images deterministically. If False, adds random crop levels; default: False
  • --deterministic_temp: Adds the deterministic temperature level to all images. If False, adds random colour temepratures in that range; default- False

  Example:

  python gen_ecg_images_from_data_batch.py -i my_input_dir -o my_output_dir --augment -rot 5 -noise 40 --deterministic_rot --deterministic_noise -se 10 --random_grid_color

  • Augmentation on synthetic images generated from the PhysioNet PTB-XL

  Adding Rotation and Crop Augmentation

  python gen_ecg_images_from_data_batch.py -i my_input_dir -o my_output_dir --augment -rot 30 -c 0.1 --deterministic_rot --deterministic_noise -se 10

• Adding all the distortions together:

  Example:

  python gen_ecg_images_from_data_batch.py -i my_input_dir -o my_output_dir --augment -rot 5 -noise 40 --deterministic_rot --deterministic_noise --hw_text -n 4 --x_offset 30 --y_offset 20 --wrinkles -ca 45 -se 10 --print_header

  • All distortions on synthetic images generated from the PhysioNet PTB-XL

  Example:

  python gen_ecg_images_from_data_batch.py -i <input_dir> -o <output_dir> --random_add_header 0.8 -se 20 --store_text_bounding_box  --resolution 300 --random_padding --pad_inches 1 --bbox --augment -rot 5 -noise 40 --deterministic_rot --deterministic_noise --hw_text -n 4 --x_offset 30 --y_offset 20 --wrinkles -ca 45 
  

  Distortionless images with Bounding box annotations

  Note: The red and green boxes here have been plotted from the corresponding text files for each lead and lead-name text.

Generating image from a single ECG record

• To provide more flexibility, we also allow users to generate synthetic ECG image from a single ECG record from a give index. The start index should be in the range: [0, len(signal))

  Example:

  python gen_ecg_image_from_data.py -i <path_to_input_file> -hea <path_to_header_file> -o <path_to_output_directory> -st start_index
  

  Note: Following pointers should be kept in mind, while generating images from a single record:

  • If the length of the signal from the -st is less than 10 seconds, the image will not be generated
  • All the batch level attributes explained above can be used to generate the image from gen_ecg_image_from_data.py

Troubleshooting

• The following command does not add handwritten text artifacts to the image:

  python gen_ecg_images_from_data_batch.py -i my_input_dir -o my_output_dir -n 4 -x_offset 30 -y_offset 20 

  Remember to enable a particular distortion to add the given artifacts.

Run-time benchmarks

Average computational time for generating an ECG image of size 2200 X 1700 pixels and 200 DPI on a MAC OS 13.4.1 (c) and Apple M2 chip

Steps	Time taken by each step per image (in seconds)
Distortion less ECG	0.72
Distortion less ECG with printed text	0.87
ECG with Hand written text distortion	6.25
ECG with Creases and Wrinkles distortions	0.92
ECG with Augmentations (Noise and rotation)	2.65
ECG with all distoritons (Hand-written text, creases, wrinkles, rotation, noise)	7.75

Citation

Please include references to the following articles in any publications:

1. Kshama Kodthalu Shivashankara, Deepanshi, Afagh Mehri Shervedani, Gari D. Clifford, Matthew A. Reyna, Reza Sameni (2024). A Synthetic Electrocardiogram (ECG) Image Generation Toolbox to Facilitate Deep Learning-Based Scanned ECG Digitization. doi: 10.48550/ARXIV.2307.01946

2. ECG-Image-Kit: A Toolkit for Synthesis, Analysis, and Digitization of Electrocardiogram Images, (2024). URL: https://github.com/alphanumericslab/ecg-image-kit

Contributors

• Kshama Kodthalu Shivashankara, School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, US
• Deepanshi, Department of Biomedical Informatics, Emory University, GA, US
• Matthew A Reyna, Department of Biomedical Informatics, Emory University, GA, US
• Gari D Clifford, Department of Biomedical Informatics, Emory University, GA, US
• Reza Sameni (contact person), Department of Biomedical Informatics, Emory University, GA, US

Contact

Please direct any inquiries, bug reports or requests for joining the team to: [email protected].