• Francois Charih (Team Leader) - [email protected] - Carleton University
• James R. Green (Principal Investigator) - [email protected] - Carleton University

This software, developed in collaboration with the Workplace Safety and Insurance Board, was designed to facilitate the digitization of audiology reports, in particular the audiograms contained therein.

This repository contains the source code for:

1. Audiogram digitization algorithm: a set of Python scripts that implement the digitization algorithm using deep convolutional neural networks.

2. Audiology report annotation interface: An electron-based desktop application that allows for the rapid annotation of audiograms contained within audiology reports. These annotations can be used to train deep convolutional neural networks to detect the elements that make up the audiogram, and that are needed for the computer to comprehend the audiogram.

• Demo: https://huggingface.co/spaces/weiren119/AudiogramDigitization

We recommend using Python 3.9 to run the algorithm as this is the version that we tested.

Before using any of this software, you must install the dependencies in a virtual environment and activate that virtual environment.

To set up the virtual environment:

$ python3 -m venv environment
$ source environment/bin/activate
$ pip3 install -r requirements.txt

This will create the virtual environment in the directory environment. This only needs to be done once. (If you are warned that the versions of torch or torchvision are not available, then, perhaps you are not using Python 3.9.)

Then, to run any command, you must ensure that this environment is activated by running (on macOS/Linux):

$ source environment/bin/activate

or on Windows

C:\> environment\Scripts\activate.bat

⚠️⚠️⚠️⚠️⚠️⚠️ Important: You will need to fix a bug in the torch library by changing F.hardswish(input, self.inplace) to F.hardswish(input) in environment/lib/python3.9/site-packages/torch/nn/modules/activation.py. ⚠️⚠️⚠️⚠️⚠️⚠️

To run the digitizer, do:

$ ./src/digitize_report.py -i <path to image or directory with images> -o <directory where the resulting JSON document should be saved>

The resulting JSON files have the same basename as the input image, but with the .json extension.

The JSON files output by the algorithm are a simple list of threshold objects that look as follows:

[
    {
        "ear": "left",
        "conduction": "air",
        "masking": false,
        "measurementType": "AIR_UNMASKED_LEFT",
        "frequency": 6000,
        "threshold": 60,
        "response": true
    },
    {
        "ear": "left",
        "conduction": "air",
        "masking": false,
        "measurementType": "AIR_UNMASKED_LEFT",
        "frequency": 3000,
        "threshold": 60,
        "response": true
    },
    ...
]

You can visually inspect the digitized audiogram as a PNG image by running the script plot_audiogram.py:

$ ./src/plot_audiogram.py -i <path_to_digitized_audiogram_json> -o <path to output image>

The models used in this algorithm are all YoloV5 models. There is an excellent tutorial on how to train a YOLOv5 model on custom data. You may download and use the annotation application to collect annotations in JSON format and use these annotations to (re)-train the object detectors.

Note that you will need to convert the locations of the objects in the JSON annotation file to the YOLO format. This is well described in the tutorial.

Once the models are trained, you can simply drop the weights (best.pt) in the appropriate location, i.e. in models/<audiograms|symbols|labels>/latest/weights, depending on whether it is the model trained to detect audiograms, symbols or axis labels.

1. Create a directory data and two subdirectories data/inputs and data/outputs.

2. Place the images in data/inputs.

3. Build the Docker container:

docker build -t digitizer .

4. Run the digitizer:

docker run -v $(pwd)/data:/app/data digitizer src/digitize_report.py -i data/inputs -o data/outputs