Specimen 01Jugend forscht state winner
PneuNet
Pneumonia detection on chest X-rays, from a NumPy CNN to a DICOM viewer.
- Period
- 2023 – 2025
- Role
- Sole author of the code; team of three at Jugend forscht
- Kind
- Research · Machine learning
- Evidence
- >87% classification · >95% segmentation
- Stack
- Python, NumPy, PyTorch, TensorFlow, OpenCV, Qt, C++
- A CNN written from scratch in NumPy, including backpropagation.
- Seven CNN architectures compared in PyTorch on four public datasets of up to 30k images, then lung segmentation and class activation maps in a Qt/C++ DICOM viewer.
- Jugend forscht 2024: Thuringian state winner and a special prize at the national final. All 60 commits are mine.
Started as a high-school research project and continued at Jugend forscht. Results are from the project's own validation split, not a clinical validation.