The task is to build a concise yet fully functional variational autoencoder that learns meaningful latent representations from X-ray images and accurately reconstructs them. I will provide an anonymised chest X-ray subset, or you may suggest a publicly available alternative that is easy to obtain and free of personal identifiers. Python 3 with PyTorch or TensorFlow/Keras is preferred. Keep the code modular, well-commented and GPU-ready so it can double as teaching material for my master’s project. The training routine should accept configurable hyper-parameters from a single file, log losses per epoch and save the best model checkpoints automatically. Deliverables • Clean source code with requirements.txt or environment.yml • Jupyter notebook demonstrating data loading, training, reconstruction and latent-space visualisation (t-SNE or PCA) • Short technical report (2-3 pages) detailing architecture, loss function, learning curves, sample reconstructions, and suggested clinical extensions • Readme explaining setup and execution Acceptance criteria • Reconstruction SSIM ≥ 0.80 on the validation set • Notebook shows at least six side-by-side original vs reconstructed X-rays • Code runs end-to-end with the provided dataset and instructions Please ensure all images remain anonymised and that no PHI is ever stored. An initial draft within one week and a final revision after feedback will keep the project on schedule. Experience with medical imaging, DICOM handling or datasets such as NIH or MIMIC-CXR is a definite plus.