🩺🧠 PulmoVision Pro — AI-Powered Pneumonia Detection

Deep Learning–based pneumonia detection from chest X-rays with Grad-CAM interpretability

🔍 Project Overview

Pneumonia is a lung infection causing inflammation in the air sacs, making early detection critical. Manual diagnosis using chest X-rays can be slow and error-prone. PulmoVision Pro automates this process using CNNs to classify Normal vs Pneumonia X-ray images, providing interpretable results with Grad-CAM heatmaps. Goals:

• Automate pneumonia detection from chest X-rays 🖼️
• Provide interpretable results for clinicians 🔥
• Deploy predictions on a professional Streamlit dashboard 💻

🎯 Objectives

• Preprocess chest X-ray images 🖼️
• Train & fine-tune CNNs (DenseNet121, ResNet50) 🧠

Evaluate models using:

• Accuracy ✅
• Precision 🎯
• Recall 📊
• AUC 📈

📊Visualize results:

• Training/Validation curves
• Confusion Matrix heatmaps
• ROC curves
• Grad-CAM overlays 🌟
• Compare DenseNet vs ResNet performance ⚖️
• Deploy predictions on interactive Streamlit dashboard

📁 Dataset

• Kermany Chest X-Ray Pneumonia Dataset (~5 GB)
• Balanced classes: Normal & Pneumonia ⚖️
• Preprocessed and ready for CNN training/testing

Kaggle link: Chest X-Ray Images (Pneumonia) Folder Structure Example:

 chest_xray/
        train/
            NORMAL/
            PNEUMONIA/
        val/
            NORMAL/
            PNEUMONIA/
        test/
            NORMAL/
            PNEUMONIA/ 

🧠 Model Architecture

Model Description

• DenseNet121 🔹 Dense connections for feature reuse; excellent for detecting X-ray textures
• ResNet50 🔹 Residual connections prevent vanishing gradients; performs well on smaller datasets

Custom Classifier:

 GlobalAveragePooling2D → Dense(128, ReLU) → Dense(1, Sigmoid) 

🛠️ Streamlit Dashboard Features

Upload single or multiple X-ray images Select DenseNet121 / ResNet50 models Display prediction label + probability Grad-CAM heatmap overlay for interpretability 🔥 Batch prediction with CSV download Interactive training/validation curves, ROC curve, confusion matrix Export predictions & heatmaps as PDF reports 📄

📊 Evaluation & Expected Results

• Model Accuracy Precision Recall AUC
• ResNet50 92–95% High High 0.95+
• DenseNet121 95–97% Very High Very High 0.97+

Notes:

• DenseNet121 generally outperforms ResNet50 due to better feature reuse
• Grad-CAM provides visual interpretability 🔥

📈 Visualizations

• Training & Validation Curves — Monitor overfitting/underfitting 📈
• Confusion Matrix Heatmap — Professional view of True vs Predicted ✅
• ROC Curve & AUC — Evaluate model performance 📊
• Grad-CAM Overlay — Highlights regions contributing most to predictions 🌟
• Sample Prediction Gallery — Multiple images with predicted labels 🖼️

🎥 Demo (GIF)

> Replace the GIF URL above with your **actual Streamlit app recording** for portfolio-ready visualization.

💻 Installation & Setup

1️⃣ Clone repository

 git clone https://github.com/ayush13-0/PulmoVision-Pro-AI-Powered-Pneumonia-Detection/tree/main
    cd PulmoVision-Pro 

2️⃣ Create virtual environment

 python -m venv venv
    # Linux/Mac
    source venv/bin/activate
    # Windows
    venv\Scripts\activate 

3️⃣ Install dependencies

 pip install -r requirements.txt 

4️⃣ Run Streamlit App

 streamlit run PulmoVision-Pro.py 

5️⃣ Download Dataset

• Kaggle Chest X-Ray Pneumonia
• Organize folder structure as shown above

🔗 Pre-trained Models

 
- DenseNet121: models/densenet_pulmovision.h5
- ResNet50: models/resnet_pulmovision.h5
- Pre-trained on Kermany X-Ray Pneumonia dataset 

:- Ready for inference & Grad-CAM visualization 🔥

🏁 Conclusion

PulmoVision Pro demonstrates:

• Automated pneumonia detection using CNNs 🧠
• Transfer learning improves medical imaging performance 🚀
• Professional evaluation using accuracy, precision, recall, F1-score, AUC 📊
• Grad-CAM visualization for clinician-friendly interpretability 🔥
• Fully professional, interactive Streamlit dashboard for deployment 🩺

📖 References

Kermany Chest X-Ray Pneumonia Dataset – Kaggle

• He, K. et al. "Deep Residual Learning for Image Recognition", 2015
• Huang, G. et al. "Densely Connected Convolutional Networks", 2017

👨‍💻 Author

Ayush

Aspiring Data Scientist & Analyst

• 📫 Email: bhanuseenu914@gmail.com
• 🌐 GitHub: https://github.com/ayush13-0
• ℹ️ LinkedIn: www.linkedin.com/in/ayush130

🛡️ License

• This project is licensed under the MIT License.