Medical Image Denoising Using Convolutional Autoencoders

This project implements the denoising autoencoder described in the paper “Medical Image Denoising Using Convolutional Denoising Autoencoders” and compares its performance against median filtering on a dental X‑ray dataset.

📁 Project Structure

medical-image-denoising/
├── notebooks/
│   └── denoising_autoencoder.ipynb   # Kaggle/Colab notebook implementing the method
├── data/
│   └── Dataset/                      # Grayscale dental X‑ray images (.jpg)
├── results/
│   ├── figures/                      # Sample denoising visualizations
│   └── metrics.txt                   # PSNR comparison results
└── README.md                         # This file

⚙️ Requirements

• Python 3.7+
• TensorFlow / Keras
• numpy, pandas
• matplotlib, OpenCV

Install via pip:

pip install tensorflow numpy pandas matplotlib opencv-python

📂 Data

The dataset consists of 120 grayscale dental X‑ray images located in data/Dataset/. Example files:

/kaggle/input/medical-image-dataset/Dataset/1.jpg
/kaggle/input/medical-image-dataset/Dataset/2.jpg
…

Images are resized to 64 × 64 pixels and normalized to [0,1].

🧠 Model Architecture

A convolutional denoising autoencoder with symmetric encoder–decoder layers:

• Encoder: two Conv2D+ReLU blocks with 64 filters, followed by MaxPooling
• Decoder: two Conv2D+ReLU blocks with UpSampling, ending in a single-channel sigmoid output
• Loss: binary crossentropy

Model summary:

Input: (64,64,1)
Conv1 → pool1 → Conv2 → pool2 → Conv3 → upsample1 → Conv4 → upsample2 → Conv5
Output: (64,64,1)

🚀 Training

1. Noise injection: Gaussian noise (σ=1) scaled by 0.07 is added to clean images.
2. Train/test split: first 100 images for training, last 20 for validation.
3. Hyperparameters: 40 epochs, batch size 10, Adam optimizer.
4. Early stopping: patience=10 on validation loss.

Notebook cell example:

model.fit(x_noisy, x_noisy,
          epochs=40,
          batch_size=10,
          validation_data=(x_test_noisy, x_test_noisy),
          callbacks=[EarlyStopping(...)] )

📊 Evaluation

• Visual comparison: plots of original, noisy, denoised (autoencoder), and median‑filtered images.

• Quantitative metric: Peak Signal‑to‑Noise Ratio (PSNR). Example:

  • Autoencoder: 69.89 dB
  • Median filter: 58.46 dB

Compute PSNR:

def PSNR(orig, denoised):
    mse = np.mean((orig - denoised)**2)
    return 20 * log10(255.0 / sqrt(mse))

▶️ Usage

1. Open denoising_autoencoder.ipynb in Kaggle or Colab.
2. Ensure the dataset is available under data/Dataset/.
3. Run all cells to preprocess, train, and evaluate.
4. View sample outputs in the results/figures/ folder.

📖 References

• The original paper: Medical Image Denoising Using Convolutional Denoising Autoencoders.
• Median filtering baseline: OpenCV cv2.medianBlur with kernel size 5.