PneumoScan AI 🩺

A GUI-based deep learning application for the automated detection of pneumonia from chest X-ray images. Designed with accessibility and clinical utility in mind, PneumoScan AI combines a user-friendly interface with a powerful stacked CNN model (MobileNetV2 + DenseNet169), achieving a classification accuracy of 92%.

Demo

---

🧪 Features

• 🧠 Stacked Deep Learning Model: Combines MobileNetV2 and DenseNet169 using transfer learning for superior image classification.
• 🖼️ Image Preprocessing: Automatic resizing, normalization, and augmentation for training robustness.
• 💻 GUI Interface: Built with Tkinter for a seamless user experience.
• 📈 Real-time Results: Upload a chest X-ray and receive a diagnosis (NORMAL or PNEUMONIA) with a confidence score.
• 📊 Evaluation Tools: Confusion matrix, classification report, and training/validation curve visualizations included.

---

📁 Project Structure

PneumoScanAI/
│
├── predictors/ # Model and prediction logic
│ ├── pneumonia_classifier.h5
│ ├── stacked_model.h5
│ └── pneumonia.py
│ └── model.ipynb # Model training notebook
│
├── main.py # Simple application
├── gui.py # GUI application
├── final_report.pdf # Project Report
└── README.md

---

🛠️ Tech Stack

• Language: Python 3.12.4
• GUI: Tkinter
• Deep Learning: TensorFlow, Keras
• Image Processing: OpenCV
• Evaluation: scikit-learn, matplotlib, seaborn

---

🧠 Model Details

• Input shape: 224x224x3
• Backbone networks: MobileNetV2, DenseNet169
• Layers: Feature maps concatenated, followed by fully connected layers with dropout
• Optimizer: Adam (lr=0.0001)
• Loss Function: Binary Crossentropy
• Accuracy: 92% on test set (624 samples)

---

How It Works

1. Launch the App (gui.py)
2. Upload a chest X-ray (JPG/PNG)
3. The system will:
   • Preprocess the image
   • Run inference using the trained stacked model
   • Display diagnosis and confidence score in GUI

---

📊 Results

• Accuracy: 92%
• Precision/Recall/F1: Evaluated via scikit-learn's classification report
• Visualizations: Training curves and confusion matrix included

---

Limitations

• Dataset imbalance between NORMAL and PNEUMONIA classes
• Binary classification only (does not detect other lung diseases)

---

Future Work

• Expand dataset and add multi-class detection
• Integrate attention mechanisms for ROI visualization
• Enable DICOM file support
• Cloud-based deployment for wider accessibility

---

Acknowledgments

Developed as part of the Artificial Intelligence course (BSCS-515) at University of Karachi (UBIT). We thank open-source contributors and platforms like Kaggle, TensorFlow, and Keras for enabling this project.

Team Members:

• Muhammad Bilal Khan (Group Leader)
• Hafiz Muhammad Shahrayar
• Haseeb Ahmed
• Muhammad Abdullah
• Muhammad Wasif Raza
• Syed Zawar Hussain

---

📄 License

This project is licensed under the MIT License.
Feel free to use, modify, and distribute it with proper attribution.