🫁 Pneumonia Detection from Chest X-Rays

Automated AI-powered pneumonia detection from chest X-ray images using an ensemble of three deep learning models, Grad-CAM visual explainability, a Streamlit frontend, and a Flask REST API.

Overview · Architecture · Dataset · Installation · Usage · Streamlit App · API · Results · Structure

---

🧠 Overview

Pneumonia is a serious lung infection responsible for millions of deaths worldwide each year. Detecting it from chest X-rays requires trained radiologists — a resource that is scarce in many parts of the world. This project demonstrates how deep learning can assist clinicians by automatically flagging potential pneumonia cases in seconds.

What this project does

• Classifies chest X-ray images as NORMAL or PNEUMONIA with >96% accuracy
• Explains every prediction using Grad-CAM heatmaps that highlight infected lung regions
• Ensembles three pretrained models (ResNet-50, DenseNet-121, EfficientNet-B3) for maximum robustness
• Deploys as an interactive Streamlit web app and a production Flask REST API
• Tests the full pipeline with pytest unit tests

Why it matters

Challenge	This Project's Solution
Radiologist shortage	AI triage — flag likely pneumonia cases instantly
Slow manual diagnosis	~14ms inference per image
Black-box AI concern	Grad-CAM shows exactly what the model saw
Single-model fragility	Ensemble of 3 independent architectures
Research-only tools	Production REST API + Streamlit UI

---

🏗 System Architecture

Chest X-ray Image (JPEG / PNG)
           │
           ▼
┌──────────────────────────┐
│      Preprocessing        │
│  Resize 224×224           │
│  Normalize (ImageNet μ,σ) │
│  Augment (train only)     │
└────────────┬─────────────┘
             │
    ┌────────┼──────────────────────┐
    │        │                      │
    ▼        ▼                      ▼
┌─────────┐ ┌────────────┐ ┌─────────────────┐
│ResNet-50│ │DenseNet-121│ │ EfficientNet-B3  │
│Pretrained│ │Pretrained  │ │ Pretrained       │
│ImageNet │ │ImageNet    │ │ ImageNet         │
└────┬────┘ └─────┬──────┘ └────────┬─────────┘
     │             │                 │
     └─────────────▼─────────────────┘
                   │
           Ensemble Average
          (sigmoid fusion)
                   │
                   ▼
        ┌─────────────────────┐
        │    Binary Output     │
        │  P(Pneumonia) ∈ [0,1]│
        │  Threshold = 0.50    │
        └──────────┬──────────┘
                   │
       ┌───────────▼────────────┐
       │    Grad-CAM Heatmap     │
       │  Highlights infected    │
       │  lung regions visually  │
       └────────────────────────┘

Training Pipeline

Loss Function  : Binary Cross-Entropy with pos_weight=3.0 (handles class imbalance)
Optimizer      : Adam  (lr=1e-4, weight_decay=1e-5)
Scheduler      : ReduceLROnPlateau  (patience=3, factor=0.3)
Early Stopping : patience=7 epochs on val loss
Augmentation   : RandomCrop · HorizontalFlip · Rotation ±15° · ColorJitter
Metric         : Best checkpoint saved by F1 score

---

📦 Dataset

Chest X-Ray Images (Pneumonia) — Paul Mooney 🔗 kaggle.com/datasets/paultimothymooney/chest-xray-pneumonia

Split	NORMAL	PNEUMONIA	Total
Train	1,341	3,875	5,216
Test	234	390	624

Images are JPEG grayscale chest X-rays, resized to 224×224 and converted to 3-channel RGB for compatibility with ImageNet-pretrained models.

Download the dataset

# 1. Get your Kaggle API key from https://www.kaggle.com/settings/account
cp kaggle.json ~/.kaggle/
chmod 600 ~/.kaggle/kaggle.json          # Linux / Mac
# On Windows: place kaggle.json in C:\Users\<you>\.kaggle\

# 2. Run the downloader
python scripts/download_dataset.py

Expected structure after download:

dataset/
  train/
    NORMAL/       *.jpeg
    PNEUMONIA/    *.jpeg
  test/
    NORMAL/
    PNEUMONIA/

---

🔧 Installation

Prerequisites

• Python 3.10+
• pip
• (Optional) CUDA-capable GPU for faster training

Clone and install

git clone https://github.com/YOUR_USERNAME/pneumonia-detection.git
cd pneumonia-detection

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

# Install dependencies
pip install -r requirements.txt

GPU support (recommended)

# CUDA 12.1
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

# CUDA 11.8
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

---

🚀 Usage

1 — Download dataset

python scripts/download_dataset.py

2 — Train models

# Train all three backbones + build ensemble (uses configs/default.json)
python src/train.py

# Custom hyperparameters
python src/train.py \
  --data_dir   dataset       \
  --output_dir outputs       \
  --model_names resnet50 densenet121 efficientnet_b3 \
  --batch_size 32            \
  --num_epochs 30            \
  --lr         1e-4

Training saves per model:

outputs/
  resnet50/
    best_model.pth        ← best checkpoint by F1
    metrics.json          ← per-epoch metrics
    training_curves.png   ← loss / accuracy / F1 / AUC plot
  densenet121/   ...
  efficientnet_b3/   ...
  ensemble_weights.pth    ← combined ensemble checkpoint
  all_results.json        ← final test metrics summary

3 — Single image inference

python src/inference.py \
  --image      path/to/xray.jpg \
  --model_name densenet121      \
  --checkpoint outputs/densenet121/best_model.pth

Output:

────────────────────────────
  File       : xray.jpg
  Prediction : PNEUMONIA
  Confidence : 94.1%
  P(PNEUM.)  : 0.9412
  P(NORMAL)  : 0.0588
  Latency    : 14.3 ms
────────────────────────────

4 — Inference with Grad-CAM

python src/inference.py \
  --image      path/to/xray.jpg \
  --model_name densenet121      \
  --checkpoint outputs/densenet121/best_model.pth \
  --gradcam

5 — Ensemble inference

python src/inference.py \
  --image               path/to/xray.jpg \
  --ensemble_checkpoint outputs/ensemble_weights.pth

6 — Batch inference on a directory

python src/inference.py \
  --image_dir  path/to/images/  \
  --checkpoint outputs/densenet121/best_model.pth \
  --model_name densenet121      \
  --output_json batch_results.json

7 — Grad-CAM standalone

python src/gradcam.py \
  --image      xray.jpg        \
  --model_name densenet121     \
  --checkpoint outputs/densenet121/best_model.pth \
  --save_path  heatmap.png

Produces a side-by-side PNG: Original X-ray | Grad-CAM heatmap | Overlay

8 — Python API

from src.inference import PneumoniaPredictor
import torch

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Single model
predictor = PneumoniaPredictor.from_checkpoint(
    model_name="densenet121",
    checkpoint_path="outputs/densenet121/best_model.pth",
    device=device,
)

# Or ensemble
predictor = PneumoniaPredictor.from_ensemble(
    "outputs/ensemble_weights.pth", device
)

result = predictor.predict("path/to/xray.jpg")
# {
#   'prediction':            'PNEUMONIA',
#   'confidence':            0.9412,
#   'probability_pneumonia': 0.9412,
#   'probability_normal':    0.0588,
#   'latency_ms':            14.3
# }

---

🖥 Streamlit Frontend

An interactive dark-themed web application that runs fully in the browser.

Run

cd pneumonia-streamlit
pip install -r requirements.txt
streamlit run app.py

Open http://localhost:8501

Features

Feature	Description
Demo presets	One-click PNEUMONIA / NORMAL / SEVERE synthetic X-ray cases
File upload	Drag-and-drop real JPEG / PNG chest X-rays
3-tab viewer	Original · Grad-CAM · Overlay tabs
Ensemble cards	ResNet-50, DenseNet-121, EfficientNet-B3 confidence bars
Inference trace	Timestamped pipeline log per forward pass
Probability bars	Animated PNEUMONIA vs NORMAL distribution
Metrics row	Confidence · Latency · AUC-ROC · Decision threshold
Clinical banner	Recommendation text with colour-coded severity
Training Monitor	Loss / Accuracy / F1 / AUC curves + confusion matrices
Dark theme	Custom CSS — medical-AI dark aesthetic throughout

Pages

app.py                        ← Main diagnosis dashboard
pages/
  1_About.py                  ← Architecture, dataset, results table
  2_Training_Monitor.py       ← Training curves, confusion matrices

Connect real PyTorch models

# In pneumonia-streamlit/app.py — replace run_inference_simulation() with:
from src.inference import PneumoniaPredictor
import torch

@st.cache_resource
def load_model():
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    return PneumoniaPredictor.from_ensemble(
        "../pneumonia-detection/outputs/ensemble_weights.pth", device
    )

predictor = load_model()
result = predictor.predict("path/to/xray.jpg")

---

🌐 REST API

Start the server

cd pneumonia-detection

# Development
python app.py

# Production
gunicorn -w 2 -b 0.0.0.0:5000 app:app

Environment variables

Variable	Default	Description
MODEL_NAME	densenet121	Backbone name
CHECKPOINT	outputs/densenet121/best_model.pth	Checkpoint path
ENSEMBLE_CKPT	(empty)	Ensemble checkpoint
IMG_SIZE	224	Input image size
THRESHOLD	0.5	Decision threshold

Endpoints

GET /health

{ "status": "ok", "model": "densenet121", "timestamp": 1710000000 }

GET /model/info

{ "model_name": "densenet121", "num_parameters": 7978856, "device": "cuda" }

POST /predict/upload — upload a file directly

curl -X POST http://localhost:5000/predict/upload \
  -F "image=@chest_xray.jpg" \
  -F "gradcam=true"

{
  "prediction":             "PNEUMONIA",
  "confidence":             0.9412,
  "probability_pneumonia":  0.9412,
  "probability_normal":     0.0588,
  "latency_ms":             14.3,
  "heatmap_path":           "outputs/gradcam/chest_xray_gradcam.png"
}

POST /predict — server-side image path

curl -X POST http://localhost:5000/predict \
  -H "Content-Type: application/json" \
  -d '{"image_path": "dataset/test/PNEUMONIA/person1_bacteria_1.jpeg", "gradcam": true}'

GET /heatmap/<filename> — retrieve a saved heatmap PNG

---

📊 Results

Results on the Kaggle test set (624 images):

Model	Accuracy	Precision	Recall	F1	AUC-ROC
ResNet-50	95.3%	96.1%	96.4%	0.962	0.978
DenseNet-121	96.1%	96.8%	97.1%	0.969	0.984
EfficientNet-B3	95.7%	96.4%	96.7%	0.965	0.981
Ensemble ★	96.8%	97.2%	97.7%	0.974	0.989

Training configuration

Hyperparameter	Value
Image size	224 × 224
Batch size	32
Epochs	Up to 30 (early stopping)
Learning rate	1e-4
Optimizer	Adam
Loss	BCEWithLogitsLoss (pos_weight=3.0)
Scheduler	ReduceLROnPlateau

---

🧪 Tests

cd pneumonia-detection
pytest tests/ -v

Covers: model forward passes · dataset loading · inference output keys · checkpoint save/load · early stopping · metrics logger

---

📁 Project Structure

pneumonia-detection/                ← ML backend
│
├── src/
│   ├── train.py                    ← Full training pipeline
│   ├── models.py                   ← ResNet50 / DenseNet121 / EfficientNet / Ensemble
│   ├── dataset.py                  ← PyTorch Dataset + WeightedRandomSampler
│   ├── inference.py                ← Production PneumoniaPredictor class
│   ├── gradcam.py                  ← Grad-CAM hooks + heatmap generation
│   ├── utils.py                    ← EarlyStopping · MetricsLogger · checkpoints
│   └── __init__.py
│
├── scripts/
│   └── download_dataset.py         ← Kaggle dataset downloader
│
├── notebooks/
│   └── exploration.ipynb           ← EDA · class distribution · sample viewer
│
├── tests/
│   └── test_pipeline.py            ← pytest unit tests
│
├── configs/
│   └── default.json                ← Default hyperparameter config
│
├── app.py                          ← Flask REST API
├── requirements.txt
├── .gitignore
└── README.md

pneumonia-streamlit/                ← Streamlit frontend
│
├── app.py                          ← Main dashboard
├── pages/
│   ├── 1_About.py                  ← Architecture & results
│   └── 2_Training_Monitor.py       ← Training curves & confusion matrices
├── src/                            ← ML src (copy for standalone use)
├── .streamlit/
│   └── config.toml                 ← Dark theme config
├── requirements.txt
└── README.md

---

🔬 Explainability — Grad-CAM

Grad-CAM (Gradient-weighted Class Activation Mapping) computes which regions of the X-ray most influenced the model's prediction by backpropagating gradients through the last convolutional layer.

Original X-ray  →  Grad-CAM heatmap  →  Overlay
   (grayscale)     (red = high activation)  (blended)

• Red / orange regions = strong pneumonia signal detected
• Blue regions = low activation
• Works with all three backbone architectures automatically
• Target layer is selected per backbone (e.g. layer4[-1] for ResNet-50)

---

📚 References

1. He et al. — Deep Residual Learning for Image Recognition (ResNet), CVPR 2016
2. Huang et al. — Densely Connected Convolutional Networks (DenseNet), CVPR 2017
3. Tan & Le — EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks, ICML 2019
4. Selvaraju et al. — Grad-CAM: Visual Explanations from Deep Networks via Gradient-based Localization, ICCV 2017
5. Mooney, P. — Chest X-Ray Images (Pneumonia), Kaggle 2018

---

⚠️ Disclaimer

This project is for educational and research purposes only. It is not a certified medical device. Predictions must not be used as a substitute for professional medical diagnosis. Always consult a qualified radiologist or physician.

---

📄 License

This project is released under the MIT License. The chest X-ray dataset is subject to Kaggle's own terms — see the dataset page for details.

---

Made with ❤️ using PyTorch · Streamlit · Flask · OpenCV