Pneumonia Detection with Deep Learning 🫁

---

Overview

This project applies Deep Learning and Computer Vision techniques to detect pneumonia from chest X-ray images.
It uses the Chest X-Ray Pneumonia Dataset from Kaggle (Kermany et al., Cell, 2018) and a DenseNet-121 architecture fine-tuned by transfer learning for binary classification (Normal vs Pneumonia).

This challenge was developed as a personal initiative, building on the knowledge gained during my Deep Learning coursework at ÉTS Montréal.

---

Dataset & Preprocessing

• Total: 5 216 training and 624 test images
• Two classes: NORMAL and PNEUMONIA
• Basic grayscale normalization and resizing to 224×224
• Optional histogram equalization and denoising

---

Model Architecture

• Backbone: DenseNet-121 pretrained on ImageNet
• Final layer: Replaced with a 2-class linear layer
• Loss: Weighted Cross-Entropy (to mitigate imbalance)
• Optimizer: AdamW (lr = 3e-4, weight_decay = 1e-4)
• Scheduler: Cosine Annealing
• Training device: Mac m1 & Ubuntu with RTX 3060 (automatic GPU detection)

---

Results

Accuracy & F1-score

• Accuracy: ≈ 93 %
• Weighted F1-score: 0.93
• NORMAL: Precision 0.98 | Recall 0.85 | F1 0.91
• PNEUMONIA: Precision 0.91 | Recall 0.99 | F1 0.95

The model achieves 93 % accuracy and a weighted F1-score of 0.93, showing a strong and balanced performance across both classes. Notably, the model maintains high precision for normal cases (reducing false alarms) while achieving near-perfect recall for pneumonia, ensuring critical cases are rarely missed.

---

Precision–Recall Curves

Both classes reach Average Precision > 0.95, showing strong separability.
The pneumonia curve stays near the top-right, while the normal class drops slightly in precision at high recall.

---

Next Steps

• Applying stronger data augmentation to the NORMAL class
• Trying Focal Loss or class reweighting
• Fine-tuning deeper layers of the DenseNet backbone

---

Run the Notebook

# Install dependencies
pip install torch torchvision torchaudio opencv-python scikit-learn matplotlib tqdm

# Launch the notebook
jupyter notebook main.ipynb

---

Author

Robert Antaluca
ÉTS Montréal / Université de Technologie de Compiègne
Website: antaluca.com

---

Sources

Pneumonia Image Classification Using DenseNet Architecture

Classification of Paediatric Pneumonia Using Modified DenseNet-121 Deep-Learning Model

Dataset

---

License

MIT License: feel free to modify and build upon this project for research or learning purposes.