Tinnitus Detection in Audiograms using Deep Learning

This repository lays the foundation for a larger vision: a human-centered, digitally supported healthcare ecosystem. We’re exploring how patients can take a more active role in their own healing process – while medical professionals regain time and clarity for what truly matters. This Project consists of a Convolutional Neural Network (CNN) developed with PyTorch to classify audiogram images as either showing normal hearing or indicating a risk of tinnitus. The model is containerized with Docker and achieves a validation accuracy of 94.58%.

---

Motivation & Problem Statement

This project holds a special, personal significance for me. After recently being diagnosed with tinnitus myself, I was motivated to apply my data science skills to better understand the condition and contribute something positive to the space.

Tinnitus is a widespread auditory condition that affects the quality of life for millions of people. Early and accurate diagnosis is crucial but can be challenging. Audiograms, which are visual representations of a person's hearing ability, often contain subtle patterns that can indicate a risk of tinnitus.

My goal was to explore whether a deep learning model could be trained to automatically recognize these patterns. In doing so, I aimed not only to deepen my own knowledge but also to create a tool that could potentially assist medical professionals in diagnosis and help raise awareness of this condition.

Understanding the Data: What is an Audiogram?

Before diving into the model, it's helpful to understand what an audiogram represents. In simple terms, an audiogram is a chart that shows the results of a hearing test. It reveals the softest sounds a person can hear at different pitches or frequencies.

• Horizontal Axis (X-axis): Represents frequency (pitch), from low pitches (like a bass drum) on the left to high pitches (like a whistle) on the right.
• Vertical Axis (Y-axis): Represents hearing level in decibels (dB), from very soft sounds at the top to very loud sounds at the bottom.

A line near the top of the chart indicates normal hearing. When the line dips downwards, it signifies hearing loss at those specific frequencies. For tinnitus-related hearing loss, it's common to see a sharp drop in the high-frequency range.

Normal Hearing vs. Tinnitus-Related Hearing Loss

Normal Hearing Example	Tinnitus-Related Hearing Loss Example

Dataset

The dataset for this project was sourced from the "Tinnitus Detection" notebook on Kaggle. Thank you to Ashik Shahriar for making this data available.

• Kaggle - Tinnitus Detection Notebook
• Kaggle - Audiological Data

The dataset consists of 1018 audiogram images. The raw data is organized into Right Ear Charts and Left Ear Charts folders. Each image is labeled by a prefix in its filename:

• N... .jpg: Normal hearing
• T... .jpg: Tinnitus diagnosed

Important Note: The raw data is not included in this GitHub repository to keep its size small. You must download the data manually from the sources linked above to run the project.

Core Methodology

• Architecture: ResNet18 with transfer learning
• Classes: normal and tinnitus (binary classification)
• Regularization: Dropout (p=0.5)
• Environment: Reproducible via Conda and Docker
• Logging: TensorBoard used to monitor training metrics

How to Run This Project

Local Training with Conda (Recommended Start)

This method creates an exact replica of the development environment on your local machine.

1. Clone the repository:

   git clone https://github.com/danielBasgo/tinnitus-trainer.git
   cd tinnitus-trainer

2. Create and activate the Conda environment: The environment.yml file contains all necessary dependencies. This command creates and activates an environment named tinnitus-trainer.

   conda env create -f environment.yml
   conda activate tinnitus-trainer

3. Download and place the data:

   • Download the data from Data Source 1. Create a folder named audiogram_dataset in the project root and place the Left Ear Charts and Right Ear Charts folders inside it.
   • Download and unzip the data from Data Source 2. Create a folder named new_dataset in the project root and place the Mild and Moderate etc. folders from the unzipped data inside it.

   Your final folder structure should look like this:

   tinnitus-trainer/
   ├── audiogram_dataset/
   │   ├── Left Ear Charts/
   │   └── Right Ear Charts/  (from Data Source 1)
   ├── new_dataset/
   │   ├──new (from Data Source 2)
   │   ├──Mild 
   │   ├──Moderate 
   ├── audiogram_dataset/      (from Data Source 1)
   │   ├── Left Ear Charts/ 
   │   └── Right Ear Charts/
   ├── new_dataset/            (from Data Source 2)
   │   └── new/
   │       ├── normal/
   │       ├── mild/
   │       └── ... (moderate, severe, etc.)
   └── ... (other files and folders)
   

4. Prepare the data: Run the script 'prepare_all_data.py' to process both datasets and create the final processed_data folder. The prepare_all_data.py script processes this raw data, splits it into a training set (80%) and a validation set (20%), and organizes it into a directory structure suitable for PyTorch's ImageFolder.

   python prepare_all_data.py

5. Train the model:

   python train.py

6. Make a prediction on a new audiogram:

After training, you can classify new images:

    python predict.py --image "path/to/your/image.jpg"

Example Output:

--- Prediction for: audiogram1.jpg ---
  -> Predicted Class: tinnitus
  -> Confidence:      58.49%

Results & Key Insights

After training on the combined dataset, the model achieved an outstanding performance with a validation accuracy of 94.58%.

A key insight came from a low-confidence prediction (58.49%) for a borderline audiogram. The model correctly identified conflicting features (normal dB levels vs. a high-frequency slope associated with tinnitus). This taught me that a robust AI model's ability to communicate its own uncertainty is a crucial feature, enabling a "human-in-the-loop" system where ambiguous cases are flagged for expert review.

Visualizing Training with TensorBoard

tensorboard --logdir=runs

Open http://localhost:6006 in your browser.

---

Training Metrics (via TensorBoard)

The training process was tracked using TensorBoard, which provides visual insights into the model's learning behavior. Below are the smoothed training and validation metrics over 10 epochs.

Accuracy	Loss

The model shows a steady improvement in both training and validation accuracy, reaching ~94.58% on the validation set. Loss curves demonstrate consistent convergence without overfitting, which indicates that the regularization (dropout) and data augmentation strategies were effective.

Docker Containerization

This project includes a Dockerfile based on the continuumio/miniconda3:latest image to provide a consistent, scientifically optimized environment. First, build the Docker image:

$ docker build -t tinnitus-trainer:latest .

Next, run a container and mount your audiogram dataset and any output directories:

$ docker run --rm \
  -v $(pwd)/audiogram_dataset:/app/audiogram_dataset \
  -v $(pwd)/new_dataset:/app/new_dataset \
  tinnitus-trainer:latest \
  python train.py

For inference with the trained model:

$ docker run --rm \
  -v $(pwd)/path/to/your/image.jpg:/app/image.jpg \
  tinnitus-trainer:latest \
  python predict.py --image "/app/image.jpg"

All dependencies are encapsulated within the container, ensuring reproducibility on any system with Docker installed.

Future Work

• ML Deployment: Deploy the trained model as a live API endpoint.
• Web App: Create a simple web interface (e.g., using Streamlit or Flask) that allows users to upload an audiogram and receive a prediction from the model's API. credit to Vivienne for building a first MVP.
• mySQL Database concept: Build a Database in mySQL credit to Janik of Team DanJanViv

---

This project was developed as part of my personal learning journey in data science and is motivated by my own experiences with the subject. A special thanks to my Teammates Vivienne and Janik (DanJanViv) for adding essential Features to this Project. And of course the tutors at DSI for their invaluable feedback and support.