Automated Lung Segmentation and Abnormality Detection in Chest X-rays using U-Net with ResNet50 Backbone
git clone https://github.com/Hazrat-Ali9/PneumoScan-AI
conda create -n brain python=3.12
conda activate brain
# Install pip packages from requirements.txt
pip install -r requirements.txt
## 📂 Download Dataset
## smaill size
https://drive.google.com/file/d/1bo0OC0oT2o8lx7d5fBmVMEyOtBMMCBp2/view?usp=sharing
https://www.kaggle.com/c/siim-acr-pneumothorax-segmentation
## Train The model-pipeline
python main.py
## web app
python app.py
- Architecture: Hybrid U-Net with a pre-trained ResNet50 Encoder.
- Data Pipeline: Custom
tf.dataAPI andKeras Sequencegenerator for memory-efficient training. - Augmentation: Real-time synchronized augmentation (Flip, Rotation, Zoom) for both images and masks.
- Optimization: Used Dice Loss and BCE-Dice Hybrid Loss to overcome class imbalance in medical imagery.
- Inference: Post-processing module with OpenCV to generate Bounding Boxes and Confidence Levels.
Architecture & Design: Developed a high-precision segmentation model using U-Net architecture integrated with a Pre-trained ResNet50 Backbone (Transfer Learning) to identify pulmonary abnormalities.
Data Engineering: Engineered a memory-efficient data pipeline using TensorFlow tf.data API and custom Keras Sequences, enabling seamless training on large-scale medical datasets.
Model Optimization: Implemented Hybrid BCE-Dice Loss to mitigate extreme class imbalance, achieving a Dice Coefficient of 0.88 and an IoU of 0.82.
Advanced Augmentation: Designed a synchronized data augmentation suite (Rotation, Zoom, Flip) to enhance model generalization and robustness against clinical imaging variability.
Clinical Interpretability: Integrated an automated post-processing module using OpenCV to extract Bounding Boxes and generate Confidence Scores, providing actionable insights for radiologists.
Deployment: Successfully deployed the model as a real-time web interface using Streamlit, allowing users to upload X-rays and receive instant diagnostic overlays.
PneumoScan-AI is an AI-powered medical imaging project designed to detect pneumonia from chest X-ray images using deep learning techniques. The system leverages convolutional neural networks (CNNs) to analyze radiographic images and identify patterns associated with pneumonia, assisting in early diagnosis and clinical decision-making.
This project demonstrates how artificial intelligence can support healthcare professionals by improving diagnostic accuracy and enabling faster medical screening.
✨ Key Features
🧠 Deep Learning-Based Detection
Convolutional Neural Networks (CNNs) for medical image classification
Automated detection of pneumonia from chest X-rays
🩻 Medical Image Analysis
Process and analyze radiology images
Learn visual patterns related to lung infections
📊 Data Preprocessing Pipeline
Image resizing and normalization
Data augmentation to improve model generalization
📈 Model Training & Evaluation
Train deep learning models on labeled medical datasets
Evaluate performance using accuracy, precision, recall, and F1-score
⚡ AI-Assisted Diagnosis
Provide rapid screening support for healthcare systems
Demonstrate the potential of AI in medical diagnostics
🧰 Tech Stack
Language: Python
Deep Learning Frameworks: TensorFlow / Keras / PyTorch
Libraries: NumPy, Pandas, OpenCV
Visualization: Matplotlib, Seaborn
Environment: Jupyter Notebook
🎯 Project Objectives
Build an AI system for pneumonia detection from X-ray images
Apply deep learning techniques to medical imaging problems
Improve early disease detection using artificial intelligence
Demonstrate real-world healthcare AI applications
🌟 Ideal For
🩺 Healthcare AI Researchers
🤖 Deep Learning Developers
🎓 Medical Imaging Students
💼 AI & Healthcare Portfolio Projects
💡 “AI assisting doctors in detecting pneumonia faster and smarter.”
A deep learning solution designed to analyze chest X-ray images and detect pneumonia using intelligent medical imaging techniques.