Lung Disease Detection from Respiratory Sounds

📌 Overview

This repository implements a hybrid deep learning framework for automatic multi-class lung disease detection from respiratory sounds.
The model integrates deep audio features (mel-spectrogram + CNN–BiLSTM–Attention) with handcrafted acoustic features (MFCCs, chroma, ZCR, spectral centroid, bandwidth).
Explainability is achieved using Grad-CAM, Integrated Gradients, and SHAP for different feature branches.

Target Diseases

• Bronchial
• Asthma
• COPD
• Healthy
• Pneumonia

🏗️ Model Architecture

The model consists of two parallel branches:

1. Mel-Spectrogram Branch

   • Input: 4s audio → Mel-Spectrogram (128 × ~250)
   • 3 Conv2D blocks with BatchNorm, ReLU, MaxPooling, Dropout
   • Flattened via TimeDistributed
   • Bidirectional LSTM (128 units × 2 directions)
   • Additive Attention → temporal context vector
   • 256-dim embedding

2. Handcrafted Feature Branch

   • Features: MFCC, Chroma, ZCR, Spectral Centroid, Bandwidth
   • Total dimension ≈ 70
   • Fully connected network (Dense(256) → Dense(128))

3. Fusion + Classification

   • Concatenate embeddings (256 + 128 = 384)
   • Dense(256) + Dropout
   • Output Softmax layer (5 classes)

⚙️ Features

• End-to-end deep + handcrafted feature fusion
• Robust data augmentations: pitch shift, time-stretch, noise injection
• Explainable AI (XAI) methods:
  • Grad-CAM on mel spectrogram
  • Integrated Gradients on mel spectrogram
  • SHAP values on handcrafted features
• Evaluation metrics:
  • Accuracy, Loss, ROC-AUC, Confusion Matrix, Classification Report
  • Per-class AUC, Micro- and Macro-averaged ROC curves

📂 Dataset

• Asthma Detection Dataset Version 2 (from Kaggle)
• Structure:

  dataset/
  ├── Bronchial/*.wav
  ├── asthma/*.wav
  ├── copd/*.wav
  ├── healthy/*.wav
  ├── pneumonia/*.wav
  

🚀 Training

• Optimizer: Adam (lr=3e-4, weight decay = 1e-4)
• Loss: Sparse Categorical Crossentropy
• Regularization: Dropout + Early Stopping
• Batch size: 16
• Epochs: 100 (with early stopping at 70)

📊 Results

• Strong validation and test accuracy across all classes
• ROC-AUC > 0.90 for all of the classes
• Grad-CAM & IG show meaningful attention on disease-relevant regions
• SHAP highlights important handcrafted features (MFCCs, spectral properties)

🔍 Explainability Examples

• Grad-CAM overlays class activation maps on mel-spectrograms
• Integrated Gradients highlights frequency bands most influential
• SHAP plots show feature importance of handcrafted features

📦 Installation

# Install dependencies (if on Colab/Kaggle, adjust as needed)
pip install numpy scipy pandas matplotlib seaborn librosa soundfile scikit-learn tensorflow==2.15.0 shap

▶️ Usage

1. Clone repo:

   git clone https://github.com/yourusername/lung-disease-detection.git
   cd lung-disease-detection

2. Prepare dataset under data_dir path inside CFG class
3. Run notebook or training script
4. Evaluate using built-in metrics
5. Visualize XAI results

📈 Visualization

• Training Curves: Accuracy & loss over epochs
• Confusion Matrix: Per-class classification performance
• ROC Curves: One-vs-rest, micro/macro average
• XAI Visualizations: Grad-CAM overlays, Integrated Gradients, SHAP barplots

🧠 Future Work

• Expand dataset with more diseases (e.g., Tuberculosis, COVID-19 coughs)
• Deploy as web app with real-time inference
• Use transformer-based encoders (AST, Wav2Vec2) for stronger embeddings

👨‍💻 Author

• S M Asiful Islam Saky
  Bachelor of Computer Science (Specialization: Data Science)
  Researcher in AI/ML/DL
  Skills: Python, TensorFlow, NLP, Data Science, Explainable AI

📜 License

This repository is licensed under the MIT License.