Core concepts

KonfAI is easiest to understand when you keep five ideas in mind:

1. YAML builds Python objects rather than acting as a loose parameter blob.

2. Datasets are organized by groups such as CT, MR, SEG, or MASK.

3. Volumes are read as patches, and a preprocessing chain that allows it is streamed from disk rather than loaded.

4. Model outputs are addressable by module path, which is how losses, metrics, and exported predictions are attached.

5. The same low-level workflow can later be packaged as a KonfAI App.

Two topics have moved out of this section: imaging-format specifics (DICOM, OME-Zarr) now live at Storage backends & formats, and packaging a finished workflow as a KonfAI App is covered in Using KonfAI Apps.

• Configuration model
• Datasets and groups
• Patch streaming
• Model graph and output naming
• Declarative YAML model graphs
• Execution flow