below is an anomaly detection framework build using streamlit. It is highly specific to my field of research (i.e., diffusion magnetic resonance imaging tractography of the brain), but thought I would share it here anyhow.
The idea behind the tool is to learn a latent representation of healthy brain features (using a deep autoencoder network) trained using healthy controls data only. Next, given a new subject (e.g., with a specific brain condition), one can look at the reconstruction error (e.g., mean absolute error) to identify potential anomalies deviating from the population norm.
For comparison, a traditional z-score based approach and a principal component analysis (PCA) + Mahalanobis distance was also implemented.
The tool also allows the visualization of “tract-profiles”, e.g., the input features.
Bonus: link to the published paper in Nature Computational Science: