Spinverse inverts face permeabilities through a differentiable Bloch-Torrey simulator to reconstruct diverse microstructural interfaces on synthetic meshes

It formulates microstructure boundary reconstruction as face-permeability inversion on a fixed grid, where low-permeability faces act as diffusion barriers, so boundaries whose topology is not fixed a priori emerge without changing mesh connectivity or vertex positions. Existing methods either assume impermeable boundaries or estimate voxel-level parameters without recovering explicit interfaces; Spinverse instead makes each interior face permeability learnable, letting the interface topology be determined by the optimization while the mesh stays unchanged. The abstract states that the method treats each interior face permeability as a learnable parameter on a fixed tetrahedral grid and recovers an interface by thresholding the learned permeability field; evidence comes from the method design and synthetic voxel mesh experiments, with no sample sizes or error values given.

By backpropagating a signal-matching loss through a fully differentiable Bloch-Torrey forward model, it can optimize face permeabilities directly from dMRI target signals. Placing dMRI inversion inside a differentiable physics simulator lets the signal-matching loss pass through the PDE forward model back to permeability parameters, rather than stopping at voxel-level parameter estimation. The abstract explicitly describes "backpropagating a signal-matching loss through the PDE forward model"; this is a method-level statement, and no specific reconstruction error metrics are reported.

It uses mesh-based geometric priors against the ill-posedness of permeability inversion and a staged multi-sequence optimization curriculum to avoid local minima. Regularization and sequence scheduling are made explicit mechanisms for ill-posed inversion and local minima, rather than relying on a single-loss optimization. The abstract notes "mesh-based geometric priors" to mitigate ill-posedness and a "staged multi-sequence optimization curriculum" to avoid local minima, and reports on synthetic voxel meshes that sequence scheduling and regularization are critical to avoid outline-only solutions.

Across a collection of synthetic voxel meshes, Spinverse reconstructs diverse geometries while improving both boundary accuracy and structural validity. Relative to methods that give only voxel-level parameters or assume impermeable boundaries, this work demonstrates recovery of explicit interfaces with structural validity on synthetic data. The abstract states "Across a collection of synthetic voxel meshes, Spinverse reconstructs diverse geometries" and that sequence scheduling and regularization improve "both boundary accuracy and structural validity"; no mesh counts, accuracy values, or baseline comparisons are given.

The work targets the inverse problem of recovering microstructural interfaces from dMRI signals, in settings where tissue is represented on a fixed tetrahedral grid and diffusion is described by a Bloch-Torrey forward model; interface resolution is limited by the resolution of the ambient mesh. The validation described in the abstract is on synthetic voxel meshes, so the results apply directly to method validation on synthetic data and provide a starting point for later evaluation under more realistic acquisition conditions. For researchers interested in differentiable physics for microstructure reconstruction, the work offers a reusable combination of face-permeability parameterization, geometric priors, and a staged multi-sequence curriculum.

The abstract does not give the number of synthetic voxel meshes, specific reconstruction error values, quantitative comparisons with existing methods, or behavior under noise, real acquisition, or different mesh resolutions. How the ill-posedness of permeability inversion varies with mesh resolution and sequence design, and how much the geometric priors and the sequence curriculum each contribute, remain open questions for a careful reader. Because the readable content here is the abstract and browse context, without body figures or experimental details, these quantitative points cannot be confirmed from the available text.