User guideΒΆ
The user guide explains the ideas shared by every model in MIR. Read Conventions first; the other pages can be read in any order.
Tensor shapes, displacement units, warping, composing flows, changing resolution, and the calling convention of each model type.
How a network output becomes a deformation: displacement, SVF, time-varying fields, symmetric mappings, LDDMM. What each guarantees.
Global smoothness penalties, spatially varying regularization (SPR) and hypernetwork-conditioned regularization strength.
Mono- and multi-modal similarity, label supervision, regularizers, Dice and Jacobian-based folding metrics.
A complete, annotated training loop with the practices used across the tutorials.
Every published checkpoint, what it was trained on, and how to load it.
Export a MIR model to ONNX and run it in a web page: the recipe behind the live spatially varying regularization playground.