Autoaudiogram: Object detection-based audiogram classification.

This repository is released as part of the publication Developing a diagnostic support system for audiogram interpretation using deep learning-based object detection. AutoAudiogram is an object detection-based audiogram classification system. The model uses object detection to detect the audiological symbols from the audiogram and then use them to classify the degree of hearing loss severity.

For the data collection, we annotation 200 audiograms including graphs, tables, and 8 audiological symbols.

Get Started

Clone the repository and navigate into the project directory

git clone https://github.com/biodatlab/autoaudiogram.git
cd autoaudiogram

Quick Started Inference

To make an inference on a given audiogram, navigate to the root of directory. Then, you can access the sample image in the sample_images directory.

from processing import feature_extraction, preprocessing_feature, classified_feature
image_dir = "sample_images/" # change this to the directory of the image
filename = "1.jpg" # change this to the filename of the image
# extract audiological symbols from image
extract_rt_df, extract_lt_df = feature_extraction(
    image_dir, filename,
    box_model_path="models/yolov8_box.pt",
    symbol_model_path="models/yolov8_symbol.pt"
)
# preprocess data for classification
feature_rt_df, feature_lt_df = preprocessing_feature(extract_rt_df, extract_lt_df)
# classify hearing severity
print(classified_feature(feature_rt_df, feature_lt_df))

To run the demo app

gradio app.py

or you can use the demo.ipynb as a starter.

Training Graph and Table Detection Model

Navigate to the directory cd src/data_extracter/graph_table and organize the folder in this directory as follows

graph_table
|____data
| |____images
| |____labels
|____runs
|____config.yaml
|____eval.ipynb
|____train.ipynb
|____train.py

Then run train.py using

python -m train

Audiological symbol training

Navigate to the directory cd src/data_extracter/symbol_detection and organize the folder as follows

symbol_detection
|____data
| |____images
| |____labels
|____runs
|____config.yaml
|____eval.ipynb
|____train.ipynb
|____train.py

Then run train.py using

python -m train

Severity classification

Navigate to the directory cd src/classification and organize folder as follows

|classification
|____data
| |____train_extract.csv # example file name
| |____test_extract.csv
| |____train_true.csv
| |____test_true.csv
|____evalute
| |____extract
| |____true
|____model
|____eval.ipynb
|____train.ipynb

python -m train --path <path to train file> # python -m train --path "data/data_extract.csv"

Evaluation

Object Detection

mAP score of each audiological symbol

| Symbols | :o: | :x: | < | > | :white_square_button: | △ | [ | ] | |---------|-----------|----------|--------------|---------------|------------------------|-----------------------|------------------------|-------------------------| | YOLOv5 | 0.957 | 0.925 | 0.848 | 0.841 | 0.833 | 0.986 | 0.826 | 0.875 | | YOLOv8 | 0.945 | 0.945 | 0.857 | 0.876 | 0.915 | 0.85 | 0.817 | 0.944 |

Hearing severity classification

Hearing loss classification performance of top-5 models using AutoML.

Automatic feature extraction | Model | Test Accuracy | Test AUC | Test Recall | Test Precision | Test F1 Score | | -------------------------------- | ----------------- | -------- | ----------- | -------------- | -------------- | | Gradient Boosting Classifier | 0.9401 | 0.9856 | 0.9401 | 0.9419 | 0.9388 | | Light Gradient Boosting Machine | 0.9472 | 0.9842 | 0.9472 | 0.9472 | 0.9472 | | Random Forest Classifier | 0.933 | 0.9806 | 0.933 | 0.9352 | 0.9317 | | Decision Tree Classifier | 0.9344 | 0.9522 | 0.9344 | 0.9346 | 0.9335 | | Extra Trees Classifier | 0.9044 | 0.9815 | 0.9044 | 0.9043 | 0.9033 |

Manual feature extraction | Model | Test Accuracy | Test AUC | Test Recall | Test Precision | Test F1 Score | |----------------------------------|-------------------|--------------|-----------------|----------------|---------------| | Gradient Boosting Classifier | 0.9629 | 0.9892 | 0.9629 | 0.9631 | 0.9629 | | Light Gradient Boosting Machine | 0.9629 | 0.9861 | 0.9629 | 0.9630 | 0.9629 | | Random Forest Classifier | 0.9643 | 0.9865 | 0.9643 | 0.9645 | 0.9643 | | Decision Tree Classifier | 0.9272 | 0.9528 | 0.9272 | 0.9286 | 0.9275 | | Extra Trees Classifier | 0.9501 | 0.9834 | 0.9501 | 0.9506 | 0.9500 |

Data Access Request

Access to our private dataset of 200 audiograms with annotation requires completion of a form. This ensures security and suitability for research/educational purposes. Submit the form here for consideration and we will grant access to the the dataset via Google Drive.

Autoaudiogram

Install / Use

README