Learning center

Multimodal fusion: alignment, source scale and display exports

Prepare fixed and moving images and understand the difference between a source display, image fusion and quantitative analysis. This guide does not demonstrate registration or agreement on an acquired study.

Prepare fixed and moving images and understand the difference between a source display, image fusion and quantitative analysis. This guide does not demonstrate registration or agreement on an acquired study.

1. Choose fixed and moving inputs explicitly

Use the intended fixed and moving images in the launcher, then open Multimodal Fusion in Image Registration. These roles are not hard-coded to MRI and PET. The current contract requires scalar intensity inputs with physical geometry; it does not manufacture a second modality from one image or treat a label map as an intensity image.

Keep with each inputReason
Role and source identityRecord which image is fixed, which is moving, and the original acquisition identity. The fixed image defines the destination grid for a resampled moving study.
Physical geometryPreserve spacing, position, orientation, coordinate convention and affine information where applicable. Equal array dimensions do not establish registration or anatomical alignment.
Axis meaningKeep spatial, time and volume-selection indices distinct. TZYX denotes a time-bearing series; a VZYX volume collection does not acquire time semantics merely because it has four axes.
Intensity meaningRetain declared pixel units and value domain. A display-normalized copy is a different representation from an input retaining its source scale; neither label supplies missing quantitative calibration.

The NIfTI-1 FAQ explains how header transformations map voxel indices into spatial coordinates and how the qform/sform codes describe those coordinate systems. Preserve that information rather than inferring orientation from the array’s shape or the way an image happens to appear on screen.

2. Review alignment before interpreting fusion

  1. Choose the Fixed spatial index, Moving spatial index, applicable time/volume indices and View spatial axis. Use APPLY INDICES / VIEW to review the actual selected inputs.
  2. Check the documented common physical grid or registration receipt. The current app refuses a pair whose only evidence is equal matrix size. Matching grid metadata is still not independent proof of correct anatomical alignment.
  3. If registration is justified, distinguish RIGID-REGISTER SELECTED 2-D PLANES from RIGID-REGISTER FULL MOVING STUDY TO FIXED 3-D. A selected-plane transform does not establish full-volume registration. These convenience actions use rigid mutual-information registration; they do not silently become affine or deformable registration.
  4. Inspect the recorded method, candidate, optimizer status, physical overlap and warnings alongside overlays or checkerboards. An optimizer converging or a similarity score improving does not independently validate the anatomical correspondence.

Retain the original inputs separately from registered derivatives. Resampling can change the grid and numerical representation even when display normalization has not been applied.

3. Read the actual dimensional scope

Execution scope offers Auto, 2-D plane, 3-D volume and 4-D frame-preserving. Inspect the actual scope and the Dimensional Capability Matrix after the run. The current source declares seventeen routes capable of volume output and three limited to planes; this declaration is not independent validation of seventeen volume algorithms.

Output or routeScope to retain
Selected QA planeA selected-plane computation that may be reduced for resource bounds. Saved 2-D route planes are not claimed to be slices of separately computed n-D route arrays.
3-D / 4-D resultSupported routes may provide a lazy volume. Frame-preserving 4-D processing applies the spatial method independently to corresponding frames; it does not perform kinetic modeling or mix information across time.
Patch / oriented-filter routesPatch Activity Selection Fusion and Oriented-Filter Fusion are explicitly 2-D approximations. Their retained route keys are sparse_representation and nsct_fusion; the implementations do not claim sparse coding or an NSCT.
Fusion Quality MetricsDescriptive plane-level Q_AB/F and Gaussian-window SSIM measures. A plane score is not a whole-volume accuracy measure, a registration certificate or modality-specific clinical quantitation.

A retained synthetic three-frame regression records 4-D frame-preserving with TZYX while keeping those three limited routes at plane_only_by_mathematical_contract. A complete run can therefore contain validly limited plane results; it does not mean that every route produced a complete 4-D array.

4. Distinguish display fusion from preserved source scale

The fusion routes work on display copies normalized to the range 0–1. For a volume plan, the current source computes one fixed minimum/maximum window per input across the frames used by the run. It does not independently renormalize each frame and erase their between-frame magnitude relationships. For a plane-only computation, the selected source planes are independently normalized.

Registered inputs retaining their source scale are kept separately from those display copies. Here, source scale preserved means that the registered/resampled input values were not passed through display normalization. It does not mean the file is byte-identical to the original acquisition, that resampling changed nothing, or that unknown units have become calibrated physical measurements.

Read the result’s intensity basis, units, dimensional scope and registration warning together. Color overlays, normalized difference maps and exploratory fusion parameter sweeps are presentation or comparison objects. The fusion routes do not calculate SUV or dynamic PET kinetic parameters; use the appropriate quantitative workflow with its own input and evidence requirements.

5. Export the representation you intend to keep

Export pathWhat to inspect
Results bundleAvailable figures, complete selected result tables, route arrays, separate source-scale input arrays and registration context. The manifest distinguishes normalized display arrays from source-scale arrays, and records whether complete 3-D/4-D outputs were exported. File counts depend on this run.
Selected table or displayThe viewer exports the selected table as CSV. Display-normalized volume results use a PNG presentation; RGB displays are not silently written as quantitative NIfTI volumes.
Source-scale scalar volumeThe viewer’s numeric path writes NIfTI plus JSON for supported scalar ZYX or explicitly temporal TZYX data. It requires suitable physical geometry and time encoding, preserves the declared float32/float64 storage type, and refuses unordered VZYX collections. Keep the sidecar with the image.
Registration and completionKeep the registration receipt and inspect whether a transform was actually exported. A failed bundle can retain a separate failure-state directory; its presence does not indicate a completed export. Check the completion state and actual manifest members.

The manifest explicitly sets independent_validation_established = false. Saving an array preserves a representation; it does not establish registration accuracy or scientific validity. This lesson checks source-defined export behavior, not a newly generated bundle or a fresh NIfTI round trip.

Worked examples describe particular inputs, methods and recorded results. Check current availability and input requirements before planning an evaluation.

← Back to the learning center