Pneumonia detection (chest X-rays)
Epitech school project: classical models and CNNs compared on chest X-rays, and what the test set revealed.
The project
Classify chest X-rays as normal or pneumonia. I compared classical models (KNN, Random Forest, SVM, with PCA for dimensionality reduction) with transfer-learning CNNs (EfficientNet, ResNet), with exploratory analysis, preprocessing and error analysis through confusion matrices.
What the test set revealed
On validation, EfficientNet reached 97% accuracy. On the official test set (624 X-rays) it dropped to 74%: it caught every pneumonia but recognised only 31% of normal X-rays. In other words, it answered “pneumonia” almost every time.
| Model | Accuracy (test) | Pneumonia recall | Normal recall |
|---|---|---|---|
| EfficientNet (validation) | 97% | 99% | 92% |
| EfficientNet (official test) | 74% | 100% | 31% |
| Random Forest (official test) | 71% | 99% | 26% |
What I took from it
A good validation metric proves nothing if the validation set doesn’t look like real data. Look at per-class recall, not just overall accuracy, and test on an independent set before drawing conclusions. It’s the same discipline I now apply to evaluating RAG systems.
Tech stack
Python
PyTorchtorchvision
scikit-learn
NumPyMatplotlibEfficientNetResNetPCA