MEDICAL-X — Chest X-Ray Disease Detection

Multi-label classification of 14 thoracic diseases from chest X-rays using a fine-tuned DenseNet121 on the NIH ChestX-ray14 dataset. Includes Grad-CAM visualization to highlight the regions the model focuses on for each predicted disease.

---

Diseases Detected

The model predicts the presence of one or more of the following conditions simultaneously:

#	Disease	#	Disease
1	Atelectasis	8	Pneumothorax
2	Cardiomegaly	9	Consolidation
3	Effusion	10	Edema
4	Infiltration	11	Emphysema
5	Mass	12	Fibrosis
6	Nodule	13	Pleural Thickening
7	Pneumonia	14	Hernia

---

Dataset

NIH ChestX-ray14 — Released by the National Institutes of Health.

• 112,120 frontal-view chest X-ray images
• 30,805 unique patients
• Labels extracted from radiology reports via NLP (weakly supervised)
• Each image can have multiple disease labels (multilabel)
• Images are 1024×1024 PNG, resized to 224×224 for training

Dataset source: NIH Clinical Center | Kaggle Download and place files in the data/ directory:

data/
├── images/
└── Data_Entry_2017.csv

---

Model Architecture

DenseNet121 pretrained on ImageNet, fine-tuned for multilabel chest X-ray classification.

DenseNet121
├── features          (pretrained, frozen backbone)
│   └── ...DenseBlocks + Transition Layers
│   └── norm5         ← Grad-CAM target layer
└── classifier
    └── Linear(1024 → 14)   ← custom head, trained from scratch

Key design decisions:

• Sigmoid activation (not softmax) — each of the 14 outputs is independent
• Binary Cross-Entropy loss — standard for multilabel classification
• One-hot encoded labels — multi-hot vectors per image
• Threshold = 0.5 — a disease is predicted positive if sigmoid output ≥ 0.5

---

Grad-CAM Explainability

Grad-CAM (Gradient-weighted Class Activation Mapping) highlights which regions of the X-ray activated the model's prediction for a given disease. The target layer is model.features.norm5 — the final batch norm of DenseNet121's feature extractor.

Input X-Ray → DenseNet121 → Sigmoid → Predicted Diseases
                  ↓
             Grad-CAM on norm5
                  ↓
         Heatmap overlaid on X-Ray

---

Project Structure

MEDICAL-X/
├── MEDICAL-X.ipynb          # Training notebook (Kaggle, GPU)
├── predict.py               # Inference script with Grad-CAM support
├── requirements.txt         # Python dependencies
├── .gitignore
└── README.md

---

Setup

1. Clone the repo

git clone https://github.com/Harsh-Prajapati54/MEDICAL-X.git
cd MEDICAL-X

2. Install dependencies

# Install PyTorch (CPU)
pip install torch torchvision --index-url https://download.pytorch.org/whl/cpu

# Install remaining dependencies
pip install -r requirements.txt

3. Download the dataset

Download the NIH ChestX-ray14 dataset from Kaggle and place it as:

data/
└── Data_Entry_2017.csv
└── images_001/images/
└── images_002/images/
...

---

Inference

Run predictions on any chest X-ray image:

# Basic prediction
python predict.py --image path/to/xray.png

# With custom model path and threshold
python predict.py --image path/to/xray.png --model densenet121_model.pth --threshold 0.4

# With Grad-CAM heatmap (saves gradcam_result.jpg)
python predict.py --image path/to/xray.png --gradcam

Example output

── Result ──────────────────────────────────────
  Image      : xray.png
  Threshold  : 0.5
  Detected   : Effusion, Infiltration

  All probabilities:
    Effusion              : 82.34% ◄
    Infiltration          : 71.12% ◄
    Consolidation         :  23.45%
    Atelectasis           :  18.90%
    ...
─────────────────────────────────────────────────

---

Training

Training was done on Kaggle (GPU T4 x2). Key hyperparameters:

Parameter	Value
Model	DenseNet121 (pretrained)
Input size	224 × 224
Batch size	32
Optimizer	Adam
Loss	Binary Cross-Entropy
Activation	Sigmoid
Classes	14

Prediction result

---

Requirements

torch
torchvision
numpy
pandas
matplotlib
seaborn
opencv-python
Pillow
tqdm

---

Acknowledgements

• NIH Clinical Center for the ChestX-ray14 dataset
• CheXNet paper — inspiration for DenseNet121 on chest X-rays
• PyTorch and torchvision for model building

---

Disclaimer

This project is for educational and research purposes only. It is not intended for clinical use or medical diagnosis.