Modeling

Modeling begins with frozen foundation-model features and ends with predictions for the research task; only the downstream path is trained.

Choose a modeling path

The structure of the encoded features and the desired prediction determine the downstream path. Each path uses the same task and training interfaces.

Feature structure	Modeling path	Typical outputs
One feature vector per sample	Apply a task-specific predictor directly.	Classification or regression; survival for slide and patient embeddings.
A bag of tile features per slide	Use Aggregators to produce one slide-level representation.	Classification, regression, or survival predictions.
A dense feature grid per tile or region	Use Decoders to recover spatial detail before prediction.	Segmentation masks or detection predictions.

Tasks and training

Patient-level pipelines use a frozen patient encoder and a task head, with no trainable aggregator.

The Tasks page defines the prediction target, loss, and compatible metrics. Training controls optimization and checkpoint selection, then executes the folds defined by the data splits. These contracts stay consistent when an aggregator, decoder, or task head is replaced.

Explore a path

The use-case table names the end-to-end scenarios that run each path on a small synthetic dataset.