Researchers have created three simulated brain phantoms to give developers a controlled reference for Susceptibility Tensor Imaging, or STI, reconstruction algorithms. Two of the phantoms were derived from STI data and the third from Diffusion Tensor Imaging, or DTI. The authors describe the set as open-source, in-silico ground-truth references, meaning computer-built versions with intended properties against which an algorithm's reconstruction can be checked. Because the phantoms are computer-built, the finding concerns benchmarking rather than clinical benefit.
STI is designed to capture susceptibility that varies with direction. The researchers used Quantitative Susceptibility Mapping, or QSM, to assess anisotropic susceptibility at different orientations. In these tests, anisotropic means the susceptibility behaves differently in different directions, a feature the reconstruction methods are expected to recover.
Inside the computer-built models
The modeled sample contained three phantoms, not three patient cases. Two were STI-derived and one was DTI-derived. The supplied report does not state how many source datasets or participants underlay those inputs. The number three describes the computer models available for testing, not a human study population or a measure of how widely the models represent brain tissue.
To build the phantoms, the team used a pipeline inspired by the QSM Reconstruction Challenge 2.0. They applied eigendecomposition to an acquired susceptibility tensor, a mathematical way of separating a directional object into its principal values and directions. The resulting mean eigenvalues, or representative strengths along those directions, were assigned to 13 distinct brain regions using values reported in the literature. This gave the models region-specific susceptibility properties rather than a single value for the whole brain.
Spatial texture was added through FA maps, which supplied the location and pattern of variation for the eigenvalues. The eigenvectors, the directional components paired with those values, came from DTI and two STI reconstructions: a Least-Squares method and Diffusion Regularized STI, or DRSTI. The construction also incorporated simulated microstructure in white-matter regions. Together, these ingredients were intended to make the phantoms resemble the structured patterns an imaging algorithm must recover.
What the tests showed
The authors ran three kinds of experiment. First, they used QSM to assess anisotropic susceptibility at different orientations. Next, they reconstructed the phantoms with five algorithms while changing the number of orientations. Finally, they repeated reconstruction across different angular rotation ranges. The design tested phantom behavior as well as the conditions under which a reconstruction algorithm might be asked to recover the underlying tensor.
In the reported visual assessment, the phantoms showed strong contrast in subcortical regions and realistic simulated microstructural patterns. Those observations speak to how the models look and behave as constructed, but the supplied analysis gives no numerical contrast score or quantitative realism measure. There is also no reported uncertainty interval or statistical test for these observations. The finding is therefore descriptive, not a measured estimate of how closely the phantoms match living human brain tissue.
Performance followed the same broad pattern across the reconstruction tests: errors decreased as the number of orientations and the rotation range increased. The authors also reported that the algorithmic trends matched prior reports. The supplied evidence does not give the size of the error reduction, identify the individual performance of the five algorithms, or specify the exact orientation counts and angular ranges. It therefore cannot show that one reconstruction method was superior to the others or quantify the benefit of a particular setting.
A benchmark, not a biological verdict
The authors concluded that all three open-source, in-silico STI brain phantoms had been developed and validated as ground-truth references for current and future reconstruction algorithms. That conclusion is strongest as a statement about computational benchmarking. Internal testing cannot establish that simulated tensors capture the full biological variation of human brain tissue, and the supplied evidence does not report the validation criteria used. Independent comparison with measured in-vivo susceptibility and microstructural data would be needed to judge realism beyond the simulations.
Open for reuse
The data, reconstruction codes and acquisition details were stated to be available through the cited Zenodo record, DOI 10.5281/zenodo.10908740, or from the authors on request. The record could help researchers reproduce the benchmark, but questions remain about how well the simulated inputs transfer to real in-vivo data and different acquisition protocols. The supplied analysis leaves those transfer questions open.
Paper data and sources
Original title: Simulation of realistic brain phantoms for susceptibility tensor imaging.
Authors: Nestor Muñoz, Carlos Milovic, Christian Langkammer, Cristian Tejos
Journal/Repository: Magma (New York, N.Y.)
Status: Peer-reviewed
First online: 2026-08-21
DOI: 10.1007/s10334-026-01402-2
Original paper · Full text