Chest_XRay_Classifier_README.md

Chest X-Ray Pathology Classifier

Automated multi-label thoracic disease detection from frontal chest X-ray images using deep learning.

Built by: Tisha Chatterjee — Biomedical Engineering Student & Independent Researcher
Research Affiliations: MIT Critical Data | Stanford | CureQuest | GeorgeWashington University |

---

What It Does

Upload a frontal chest X-ray image. Get independent probability scores for 14 thoracic pathologies plus a "No Finding" classification.

Use case: AI-assisted screening for resource-limited settings where radiologist density is critically low.

---

Dataset

Attribute	Detail
Source	NIH Clinical Center — ChestX-ray14
Total Images	112,120 frontal-view chest X-rays
Total Patients	30,805
Labels	14 thoracic diseases + "No Finding"
Label Generation	Extracted from radiology reports using Natural Language Processing (NLP)
Image Type	Frontal-view chest X-rays (PA/AP)

Classes Detected (14 + 1)

#	Disease	#	Disease
1	Atelectasis	8	Hernia
2	Cardiomegaly	9	Infiltration
3	Consolidation	10	Mass
4	Edema	11	Nodule
5	Effusion	12	Pleural Thickening
6	Emphysema	13	Pneumonia
7	Fibrosis	14	Pneumothorax
—	No Finding	—	—

Training Subset

Due to Google Colab storage constraints (~12 GB GPU RAM, limited disk), the model trains on:

Attribute	Detail
Subset	Batch 1 of NIH ChestX-ray14
Images	~4,999 (~5,000) X-rays
Size	~1 GB
Reason	Full dataset = ~45 GB — impractical for standard Colab session

Data Split

Split	Percentage	Purpose
Training	80%	Model learning
Validation	10%	Hyperparameter tuning & early stopping
Testing	10%	Final performance evaluation

---

Model Architecture

Attribute	Detail
Architecture	DenseNet-121
Framework	PyTorch
Pretrained Weights	ImageNet (14M+ natural images)
Technique	Transfer Learning → Progressive Unfreezing

Transfer Learning Strategy

1. Phase 1 — Feature Extraction:

   • Freeze all convolutional layers
   • Replace final classifier layer with 14-output linear layer
   • Train only the new classifier head
   • Epochs: 5 | LR: 0.001

2. Phase 2 — Fine-Tuning:

   • Unfreeze entire network
   • Train at reduced learning rate to preserve pretrained features
   • Epochs: 3 | LR: 0.0001

Total Training: 8 epochs

Why DenseNet-121?

• Parameter efficiency: ~8M parameters vs. ResNet-50's ~25M — faster training, less overfitting on limited data
• Feature reuse: Dense connections improve gradient flow for medical imaging fine-grained patterns
• Medical imaging standard: Widely validated on ChestX-ray14 benchmark

---

Image Preprocessing Pipeline

Step	Transformation	Detail
1	Grayscale → RGB	1-channel to 3-channel for ImageNet compatibility
2	Resize	256 × 256 pixels
3	Center Crop	224 × 224 pixels (DenseNet-121 input size)
4	Tensor Conversion	PyTorch tensor format
5	Normalization	ImageNet mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225]

Training Augmentation (DataLoader)

Augmentation	Parameters	Purpose
Random Crop	224 × 224 from 256 × 256	Spatial variability, reduce overfitting
Random Horizontal Flip	p = 0.5	Left-right symmetry in chest anatomy

Note: Aggressive augmentation avoided for medical imaging to preserve pathological features.

---

Training Configuration

Hyperparameter	Value	Rationale
Loss Function	BCEWithLogitsLoss	Multi-label classification — one X-ray may contain multiple diseases simultaneously
Optimizer	Adam	Adaptive learning rates, robust for transfer learning
Initial LR	0.001	Standard for classifier head training
Fine-Tune LR	0.0001	10× reduction to prevent catastrophic forgetting of pretrained features
LR Scheduler	ReduceLROnPlateau	Automatically reduces LR when validation loss plateaus — prevents manual tuning
Batch Size	32	Balances GPU memory utilization and gradient stability
Epochs (Total)	8	5 (feature extraction) + 3 (fine-tuning)

---

Evaluation Metrics

Metric	Purpose
ROC-AUC	Primary — threshold-independent discrimination ability per class
Precision	Of predicted positives, how many are true positives?
Recall	Of actual positives, how many did we catch?
F1-Score	Harmonic mean of precision & recall — balanced performance
Validation Loss	Model generalization monitoring
Training Loss	Convergence monitoring

Multi-label approach: Each disease predicted independently via sigmoid activation — no mutual exclusivity assumption (clinically realistic: a patient can have pneumonia + effusion).

---

Prediction Output

For each uploaded chest X-ray, the model outputs:

{
  "Atelectasis": 0.82,
  "Cardiomegaly": 0.15,
  "Consolidation": 0.91,
  "Edema": 0.07,
  "Effusion": 0.88,
  "Emphysema": 0.03,
  "Fibrosis": 0.12,
  "Hernia": 0.01,
  "Infiltration": 0.76,
  "Mass": 0.34,
  "Nodule": 0.22,
  "Pleural Thickening": 0.09,
  "Pneumonia": 0.94,
  "Pneumothorax": 0.05,
  "No Finding": 0.02
}

Interpretation: Probabilities are independent per class. A threshold (typically 0.5) converts to binary predictions. Clinical decision support, not diagnosis.

---

Technologies Used

Category	Tools
Language	Python 3
Environment	Google Colab (GPU T4)
Deep Learning	PyTorch, Torchvision
Data Processing	NumPy, Pandas, Pillow (PIL)
Visualization	Matplotlib
Metrics	Scikit-learn
Data Source	Kaggle API (NIH ChestX-ray14)

---

How to Run

Option 1: Google Colab (Recommended)

1. Click the "Open in Colab" badge above
2. Runtime → Change runtime type → Select GPU (T4)
3. Run all cells sequentially
4. Upload a chest X-ray image (PNG/JPG)
5. View predicted pathology probabilities

Option 2: Local Machine

# Clone repository
git clone https://github.com/YOUR-USERNAME/YOUR-REPO.git
cd YOUR-REPO

# Install dependencies
pip install -r requirements.txt

# Run notebook
jupyter notebook chest_xray_classifier.ipynb

Requirements:

• Python 3.8+
• CUDA-capable GPU (recommended) or CPU
• ~2 GB RAM minimum

---

Files

File	Description
chest_xray_classifier.ipynb	Main training & inference notebook
requirements.txt	Python dependencies
sample_predictions/	Example model outputs
README.md	This file

---

Limitations & Future Work

Current Limitation	Planned Improvement
Trained on ~5,000 images (Batch 1 only)	Scale to full 112,120-image dataset via cloud storage
Single frontal view	Add lateral view support for comprehensive assessment
No localization (where is the pathology?)	Integrate Grad-CAM for explainable heatmap visualization
No clinical metadata (age, sex, symptoms)	Add multimodal fusion for personalized risk scoring
No web interface	Deploy as Gradio/Streamlit app for clinician access
No regulatory validation	Pursue IRB-approved retrospective validation on hospital data

---

Clinical Disclaimer

This model is for research and educational purposes only. It is not FDA-approved, not CE-marked, and not validated for clinical diagnosis. Predictions should never replace qualified radiologist interpretation. Always consult a medical professional for diagnostic decisions.

---

About the Author

Tisha Chatterjee is a Biomedical Engineering student and independent researcher converging on brain-computer interfaces, neuroplasticity, and medical imaging. She is a researcher at MIT Critical Data, a Global Neurosurgery Fellow at George Washington University, a cohort member at Stanford-affiliated S.Y.A.L.I.S Labs (top 8% of 2,000+ applicants), and a bioinformatics researcher at CureQuest. She is a TEDx Scholar (Top 10 globally) and an IIT Bombay Techfest College Ambassador (Rank 36 / 10,000+).

• 🔗 LinkedIn
• 📧 tishachatterjee77@gmail.com
• 🧠 Research Interests: BCI · Neuroplasticity · Medical Imaging · Computational Biology · AI in Healthcare

---

Citation

If you use this work in your research, please cite:

@misc{chatterjee2026chestxray,
  author = {Chatterjee, Tisha},
  title = {Chest X-Ray Pathology Classifier: Multi-Label Thoracic Disease Detection with DenseNet-121},
  year = {2026},
  howpublished = {\url{[https://github.com/tishact7/Chest_XRay_Pathology_Classifier]}},
  note = {Independent research project}
}

---

License

Not for clinical deployment without regulatory approval (FDA, CE, CDSCO, etc.).

---

Acknowledgments

• NIH Clinical Center for the ChestX-ray14 dataset
• PyTorch Team for the DenseNet-121 implementation
• Google Colab for free GPU access enabling independent research

---

Last updated: July 2026 | Model version: 1.0 | Next release: v1.1 (Grad-CAM + Gradio UI)