Paper deep dive
$\texttt{DisMorph}$: learning to disentangle technical distortions from true biological change
Jingru Fu, Kathleen E. Larson, Douglas N. Greve, Bruce Fischl, Malte Hoffmann
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 8/11/2026, 5:28:24 AM
Summary
The paper introduces DisMorph, a deep learning-based registration framework designed to disentangle technical MRI distortions (specifically Gradient Non-Linearity) from true biological anatomical changes in longitudinal MRI scans. Trained entirely on synthetic data using a generative model that simulates both distortion and atrophy, DisMorph predicts two separate dense deformation fields: one for technical distortion and one for anatomical change. Evaluations on simulated data, fBIRN distortion-only pairs, and longitudinal Alzheimer's Disease (ADNI) data demonstrate that DisMorph outperforms conventional registration methods by reducing estimation bias and accurately isolating residual scanner-induced distortions from genuine neurodegenerative atrophy.
Entities (8)
Relation Signals (8)
DisMorph → disentangles → technical distortions
confidence 95% · DisMorph explicitly decomposes longitudinal deformation into technical and anatomical transforms.
DisMorph → disentangles → biological change
confidence 95% · DisMorph explicitly decomposes longitudinal deformation into technical and anatomical transforms.
DisMorph → detects → Atrophy
confidence 92% · On longitudinal Alzheimer’s disease (AD) pairs, our method detects anatomical change in AD-related brain structures...
DisMorph → trainedon → Synthetic Data
confidence 92% · We propose DisMorph, a registration framework trained entirely on synthetic data...
Gradient Non-Linearity → causes → geometric distortion
confidence 90% · MRI scanners introduce geometric distortions that vary across acquisition systems and protocols, such as gradient non-linearity (GNL) distortion.
DisMorph → evaluatedon → fBIRN
confidence 90% · We evaluate DisMorph on two real datasets: fBIRN [31] and ADNI [28].
DisMorph → evaluatedon → ADNI
confidence 90% · We evaluate DisMorph on two real datasets: fBIRN [31] and ADNI [28].
DisMorph → outperforms → conventional registration
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Longitudinal MRI enables sensitive measurement of structural brain change for studying aging and neurodegenerative disease. Deformable image registration is a key tool for estimating such change by computing a dense deformation that captures geometric differences between longitudinal scans. However, MRI scanners introduce geometric distortions that vary across acquisition systems and protocols, such as gradient non-linearity (GNL) distortion. Existing registration methods estimate a single field that conflates biological and technical effects, potentially biasing downstream morphometric measurements if distortions remain (partially) uncorrected. We propose $\texttt{DisMorph}$, a registration framework trained entirely on synthetic data that explicitly decomposes longitudinal deformation into technical and anatomical transforms. It predicts two dense deformations, each encoding one effect. During training, a novel generative model synthesizes both effects separately to provide disentanglement supervision, while domain randomization promotes generalization across imaging protocols. We evaluate our method in three complementary settings. On simulated data with known ground truth, our method detects anatomical change more accurately and consistently than conventional registration. On real image pairs that differ only by GNL distortion, our method assigns most geometric change to the distortion field, demonstrating specificity in the absence of anatomical change. On longitudinal Alzheimer's disease (AD) pairs, our method detects anatomical change in AD-related brain structures while identifying residual distortion left after standard correction. By disentangling MRI-induced distortion from biological change in the longitudinal deformation, our method paves the way for more accurate longitudinal morphometry in clinical settings where maintaining acquisition consistency is challenging.
Tags
Links
- Source: https://arxiv.org/abs/2608.08173v1
- Canonical: https://arxiv.org/abs/2608.08173v1
Trouble viewing inline? Open PDF directly →
Full Text
29,855 characters extracted from source content.
Expand or collapse full text
11institutetext: Athinoula A. Martinos Center for Biomedical Imaging, Charlestown, USA 22institutetext: Department of Radiology, Massachusetts General Hospital, Boston, USA 33institutetext: Department of Radiology, Harvard Medical School, Boston, USA 33email: jifu1@mgh.harvard.edu DisMorph: learning to disentangle technical distortions from true biological change Jingru Fu Corresponding author. Kathleen E. Larson Douglas N. Greve Bruce Fischl,† Malte Hoffmann,† Abstract Longitudinal MRI enables sensitive measurement of structural brain change for studying aging and neurodegenerative disease. Deformable image registration is a key tool for estimating such change by computing a dense deformation that captures geometric differences between longitudinal scans. However, MRI scanners introduce geometric distortions that vary across acquisition systems and protocols, such as gradient non-linearity (GNL) distortion. Existing registration methods estimate a single field that conflates biological and technical effects, potentially biasing downstream morphometric measurements if distortions remain (partially) uncorrected. We propose DisMorph, a registration framework trained entirely on synthetic data that explicitly decomposes longitudinal deformation into technical and anatomical transforms. DisMorph predicts two dense deformations, each encoding one effect. During training, a novel generative model synthesizes both effects separately to provide disentanglement supervision, while domain randomization promotes generalization across imaging protocols. We evaluate DisMorph in three complementary settings. On simulated data with known ground truth, DisMorph detects anatomical change more accurately and consistently than conventional registration. On real image pairs that differ only by GNL distortion, DisMorph assigns most geometric change to the distortion field, demonstrating specificity in the absence of anatomical change. On longitudinal Alzheimer’s disease (AD) pairs, DisMorph detects anatomical change in AD-related brain structures while identifying residual distortion left after standard correction. By disentangling MRI-induced distortion from biological change in the longitudinal deformation, DisMorph paves the way for more accurate longitudinal morphometry in clinical settings where maintaining acquisition consistency is challenging. $ $$ $footnotetext: Shared senior authorship. 1 Introduction Figure 1: Tissue boundary shifts induced by gradient non-linearity in MRI. (A) Sagittal fBIRN source slice [31]. (B) Pial (red) and white-matter (yellow) surfaces fitted to the distortion-corrected image. (C) The same surface overlays on the original, distorted image. Arrows highlight boundary displacements. Morphometric analysis of brain magnetic resonance imaging (MRI) enables in-vivo quantification of structural neuroanatomical properties indicative of neurodevelopment, neurodegenerative disease, and aging trajectories [4, 17, 14]. Compared to cross-sectional imaging, longitudinal imaging improves sensitivity by eliminating inter-subject anatomical variability and enables within-subject comparisons over time [38, 36, 37]. Accurate estimation of longitudinal change is therefore fundamental to brain morphometry and disease progression studies. Unfortunately, imperfections in MRI hardware introduce undesirable, non-anatomical distortions. Differences between imaging systems and acquisition protocols cause these distortions to vary across longitudinal scans, which can obscure true biological change, thereby reducing statistical power and biasing estimates of disease progression [30, 43, 39]. Spatial distortions are particularly problematic for morphometry because they directly affect measurements of volume [12], shape [34], and tissue boundaries [8]. Gradient non-linearity (GNL) is the most prominent source of distortion in structural MRI (Fig. 1) [30]. Although GNL correction is available on modern MRI systems, this correction is not always applied and can be incomplete: residual distortion can persist even after vendor correction, with reported magnitudes of up to 2 m within a typical brain-imaging field of view, comparable in scale to annual atrophy rates measured in neurodegeneration studies [2, 35]. Furthermore, retrospective datasets often provide no reliable way of verifying whether studies applied GNL correction correctly or at all. For instance, ADNI phantom monitoring identified issues including GNL correction with coefficients for the wrong gradient system, disabled autoshimming, and a miscalibrated laser alignment light, which would have compromised morphometry at over 25% of participating sites had they remained undetected [20]. As an alternative to vendor correction, phantom-based calibration compares image geometry to a known phantom geometry. Although useful, it is impractical for retrospective, legacy, or large multi-site studies. Therefore, image-based approaches capable of separating residual MRI-induced distortion from biological change in longitudinal image pairs would add value and complement vendor-provided GNL correction. A core technique for quantifying longitudinal change is deformable image registration, which estimates a dense deformation field that captures geometric differences between scans. Classical registration has undergone decades of development [6, 5, 41, 44]. More recently, deep-learning methods have substantially improved computational efficiency and registration accuracy [11, 10, 7, 21, 18, 42]. Among these, several works [21, 24, 9, 26, 22, 23, 1] use highly variable synthetic training images to enhance robustness to image contrasts, resolutions, and acquisition conditions. However, most learning approaches focus exclusively on inter-subject alignment rather than longitudinal change [15]. More importantly, both classical and learning-based registration methods aim to explain all spatial differences in a single deformation, without the ability to disentangle technical distortions from true biological change. In this study, we introduce DisMorph, a learning-based registration framework that explicitly disentangles MRI distortions from biological change. It uses a novel generative strategy to synthesize diverse spatial effects—GNL distortion and atrophy—together with intensity variations that make the registration robust to real-world variability in acquisition hardware and protocols. We demonstrate in controlled simulations that DisMorph recovers the underlying sources of three-dimensional (3D) geometric effects, and show on real datasets that its decompositions are consistent with the expected technical and anatomical differences, enabling reduction of distortion-related bias for downstream analysis. Figure 2: DisMorph strategy. We synthesize longitudinal training pairs with a geometry generator that separately applies distortion and atrophy to label maps, followed by an intensity generator that synthesizes images with corruptions from them. A 3D registration model learns to disentangle both geometric effects. 2 Methods The proposed framework, DisMorph, combines a procedural, generative model that synthesizes longitudinal image pairs for training and a registration network capable of disentangling these effects by predicting separate deformation fields for distortion and atrophy. Fig. 2 provides an overview of our strategy. 2.1 Synthesizing Longitudinal Training Data DisMorph synthesizes images of diverse geometry and contrast from anatomical label maps using two procedural models: (i) a geometry generator that spatially augments label maps and (i) an intensity generator that creates variable gray-scale images from them. These models synthesize diverse deformations, image contrasts, and corruptions on the fly. Fig. 3 illustrates the data generation. Geometry generator. Let L denote a multi-class anatomical label map sampled from a training dataset =L(n)n=1ND=\L^(n)\_n=1^N, delineating K distinct brain structures. We spatially augment L by applying a nonlinear spatial transform. To simulate compound longitudinal effects, we employ two additional spatial transformations. First, we apply a smooth random elastic transform to L that models GNL distortion, yielding the distorted label map LdL_d. We generate this transform by sampling displacement vectors on a coarse control lattice and interpolating it to full resolution. We sample vector magnitudes up to 2.5% of the field of view and lattice resolutions between 2 and 10 control points for each dimension, varying the smoothness of the vector field. We integrate the resulting field using scaling and squaring (S) [3] to obtain a diffeomorphism. This stochastic generation produces diverse distortions. Next, we introduce anatomical changes by moving label boundaries in LdL_d. Specifically, we choose a label subset ⊂1,…,KK⊂\1,…,K\ of brain regions and simulate local atrophy or growth through iterative boundary reassignment between each target structure and its neighboring labels. For each label k∈k , we uniformly sample a volumetric change ratio rk∼(−0.5, 0.5)r_k (-0.5,\,0.5), defined as rk=Vktarget/Vkorig−1r_k=V_k^target/V_k^orig-1, where negative values simulate atrophy, and positive values simulate growth. We use distance maps to identify boundary voxels shared with adjacent structures and reassign these voxels to progressively shrink or expand the target region until the resulting volume change matches or exceeds rkr_k. Adding distortion and tissue change yields the final label map LdaL_da. Intensity generator. We generate intensity images (I,Ida)(I,I_da) from label maps (L,Lda)(L,L_da) following prior synthesis-based approaches [9, 22, 15]. For each image I, briefly, we assign a different, random mean intensity μk _k to voxels x associated with each label k and sample intensities normally, I(x)∼(μk,σk2)I(x) ( _k, _k^2), where μk∼(0,1) _k (0,1) and σk∼(0,0.05) _k (0,0.05). Next, we apply a set of image corruptions, including bias fields, gamma transforms, blurring, and additive noise [9, 15]. Figure 3: Example of randomized synthesis steps (Fig. 2). The realization shows hippocampal atrophy and ventricular enlargement across both hemispheres. 2.2 Disentangling Distortion and Atrophy We train a registration model using synthetic intensity images as input and supervise it with the corresponding label maps. This strategy removes dependence on specific image contrast and encourages the model to focus on anatomy and geometry instead. After training, the model can register images without requiring label maps. Let IFI_ F and IMI_ M denote a fixed and a moving image, respectively, with corresponding label maps LFL_ F and LML_ M. The network predicts two stationary velocity fields (SVFs), one capturing MRI distortions (vdv_d) and the other capturing anatomical change (vav_a). To obtain diffeomorphisms, we integrate each SVF via S, yielding deformation fields ϕd=exp(vd) _d= (v_d) and ϕa=exp(va) _a= (v_a) that warp the moving image towards the fixed image. We represent label maps using one-hot encoding, denoted by L~ L, and train the model by optimising the mean squared error between label representations. Given a deformation ϕφ, we define the loss as ℒlabel(L~M,L~F;ϕ)=1|Ω|∑x∈Ω‖(L~M∘ϕ)(x)−L~F(x)‖22,L_label( L_ M, L_ F;φ)= 1| | _x∈ \|( L_ M φ)(x)- L_ F(x) \|_2^2, (1) where Ω denotes the spatial domain and L~M∘ϕ L_ M φ is L~M L_ M deformed by ϕφ. We use separate distortion and anatomical terms ℒd=ℒlabel(L~,L~d;ϕd)L_d=L_label( L, L_d; _d) and ℒa=ℒlabel(L~,L~da;ϕda)L_a=L_label( L, L_da; _da), respectively, and denote the composed deformation by ϕda=ϕd∘ϕa _da= _d _a. Let ϕda(x)=x+u(x) _da(x)=x+u(x) denote the displacement formulation of the deformation. We encourage warp smoothness by adding regularization term ℒsmooth=‖∇u‖22L_smooth=\|∇ u\|_2^2, yielding the final training objective ℒtotal=ℒd+ℒa+λℒsmooth,L_total=L_d+L_a+λ\,L_smooth, (2) where λ controls the regularization strength. 3 Experiments and Results We design our experiments to address two key questions. (1) To what extent can DisMorph reduce errors in the estimation of known atrophy compared to conventional registration in distorted image pairs? (2) Are the two transforms predicted by DisMorph sufficiently specific to disentangle GNL distortion from anatomical atrophy in real MRI data? We address (1) in Sec. 3.1 using simulated longitudinal MRI generated from a biomechanical atrophy model with known ground-truth tissue change. We compare the change estimated by DisMorph with the known ground truth. We address (2) in Sec. 3.2 using two complementary real MRI datasets. First, we evaluate performance on image pairs that differ only by GNL distortions. Second, we evaluate performance on longitudinal image pairs from subjects with Alzheimer’s disease (AD), acquired two years apart and expected to differ primarily due to atrophy. Training Data. We train DisMorph on images synthesized from 530 label maps obtained by segmenting structural, T1-weighted images from IXI [27] and FSM [19] using FreeSurfer [13, 37, 25]. We affine-align all subjects to Talairach space and crop images to a fixed size of 160×192×224160× 192× 224 isotropic 1-m voxels. Metrics. For each structure s, we compute the absolute symmetrized percent change (ASPC) [37] between the moving and warped label maps. Specifically, we obtain sub-voxel volume estimates by interpreting interpolated one-hot values as partial-volume fractions. We compute ASPC as: ASPCs=100×|Vmoving,s−Vmoved,s|0.5(Vmoving,s+Vmoved,s). ASPC_s=100× |V_moving,s-V_moved,s|0.5\,(V_moving,s+V_moved,s). (3) Implementation Details. As a general registration framework, DisMorph can adopt any backbone to estimate deformation fields. We choose a dual-stream encoder with shared weights and filter sizes 32, 64, 64, 64 to encode fixed and moving images separately. We then use a pyramid-warping encoder [29] with filter sizes 128, 128, 128, 3 to estimate deformation fields in a coarse-to-fine manner. The network predicts the distortion field at quarter resolution and the anatomical field at half resolution. We upsample both fields to full resolution to compute the losses (Sec. 2.2). Model training uses the Adam optimizer with batch size 1 and learning rate 10−410^-4 for 2,600 epochs with 200 steps per epoch, after which the loss visually converges. We set the smoothness regularization weight to λ=0.01λ=0.01 based on empirical tuning. 3.1 Experiments on Simulated Data Figure 4: Atrophy detection error (AAE) on simulated data across three atrophy levels. Lower values indicate better agreement with the ground truth. Simulated Data. We use 150 subjects from ADNI 1 (1.5T and 3T) [28] and MIRIAD [33], divided equally into low-, moderate-, and high-atrophy groups. For each subject, we prescribe annual volume change rates for three AD-affected regions—ventricle, hippocampus, and amygdala—based on prior studies [40, 16]. Low: ventricles (2±1)%(2± 1)\%, hippocampi (−1±0.5)%(-1± 0.5)\%, amygdalae (−1±0.5)%(-1± 0.5)\%. Moderate: ventricles (3±1.5)%(3± 1.5)\%, hippocampi (−1.5±1)%(-1.5± 1)\%, amygdalae (−1.5±1)%(-1.5± 1)\%. High: ventricles (4.5±2)%(4.5± 2)\%, hippocampi (−2±1.5)%(-2± 1.5)\%, amygdalae (−2±1.5)%(-2± 1.5)\%. We simulate atrophy using a biophysical model [32], then add distortion to create distorted longitudinal pairs. Setup. We measure atrophy estimation bias as the absolute ASPC error (AAE) between the estimated and prescribed ASPC. As a baseline (Naive), we compute ASPC from the composite deformation field, which includes both anatomical change and MRI-induced distortion. For DisMorph, we instead compute ASPC from the estimated atrophy field alone. This comparison measures the benefit of disentangling MRI-induced distortion from true anatomical change. Results and Discussion. Fig. 4 shows that DisMorph consistently reduces AAE relative to Naive across groups, demonstrating robustness to atrophy magnitude and location. Importantly, DisMorph performs best where it matters most: low atrophy corresponds to early disease stages, where distortion removal yields the largest reduction in estimation bias [38]. 3.2 Experiments on Real-World Data A fBIRN (18 distortion-only pairs) B ADNI (40 longitudinal AD pairs) Figure 5: Volume change (ASPC) in real data, captured by Naive, which includes distortion and atrophy; and separate DisMorphϕd DisMorph_ _d and DisMorphϕa DisMorph_ _a fields. Figure 6: Longitudinal registration examples. We overlay target-image label contours on all images (hippocampus in yellow; gray matter in red) and highlight changes with arrows. Left: DisMorph detects distortion. The example also shows a skull-stripping inconsistency between the two images, which DisMorph interprets as apparent atrophy. Right: DisMorph captures subtle MRI-induced misalignment in the distortion field (row 2) while preserving hippocampal atrophy in the atrophy field (row 3). Data. We evaluate DisMorph on two real datasets: fBIRN [31] and ADNI [28]. First, we select one scan from each of 18 fBIRN subjects that we distortion-correct with mri_gradunwarp using the manufacturer’s coefficient files [30]. We pair up the original and gradient-corrected images; their geometric differences arise only from distortion (Fig. 1). Second, we select longitudinal image pairs acquired two years apart from 40 ADNI 1 (1.5T) subjects diagnosed with AD. These images have been distortion-corrected in preprocessing. Although some residual, differential distortion typically remains, we expect the dominant geometric effect to be atrophy. Setup. We compute per-structure ASPC between moving label maps and those moved by Naive registration ϕφ, the DisMorph distortion field ϕd _d only, and the DisMorph anatomical field ϕa _a after removing the distortion component. Results and Discussion. Fig. 5A reports ASPC measured from the three deformation fields for fBIRN [31] and Fig. 5B for ADNI [28] across subcortical structures. For the fBIRN pairs, DisMorph correctly attributes the majority of geometric differences to distortion. The residual signal in the atrophy field is minimal and might arise from preprocessing errors. We show an example in Fig. 6 (left), in which the distortion field (row 2) explains nearly all geometric differences, such that the source image deformed by the distortion field aligns closely with the target image, leaving an almost zero atrophy field. In contrast, row 3 highlights an inconsistency in skull-stripping, where tissue retained in the target image is absent from the source image. The model assigns this localized tissue loss to the atrophy field, likely because it resembles anatomical change rather than technical distortion. For AD pairs, Fig. 5B shows that the atrophy field captures the bulk subcortical volume change, consistent with the expected disease trajectories [40, 16]. The example in Fig. 6 (right) contains a large atrophy component and a small distortion component. The hippocampal volume loss visible after two years (row 3) remains preserved after applying the estimated distortion field, indicating that DisMorph separates the smooth MRI-induced distortion from localized tissue change. Although trained solely on synthetic data, DisMorph generalizes well to both real datasets and produces a biologically plausible decomposition of technical and anatomical deformation. 4 Conclusion We introduce DisMorph, the first registration framework that explicitly decomposes longitudinal deformation into technical and anatomical transforms. Across simulated and real datasets, our results demonstrate that this decomposition improves estimation of tissue change while isolating technical deformation. The ADNI analysis suggests that residual scanner-induced distortion persists despite standard vendor correction. The present study evaluates the model’s ability to isolate gradient nonlinearity. Future work will extend the evaluation to additional distortion sources, including B0 field inhomogeneity and chemical shift [43]. We also plan to investigate the impact of deformation decomposition on downstream morphometric analyses and statistical models of disease progression. Explicitly simulating the physical origin of deformation in training may provide value to longitudinal morphometry in heterogeneous, retrospective, and multi-site imaging studies by complementing vendor corrections. credits 4.0.1 Acknowledgements JF was supported by a WASP International Postdoctoral Scholarship from the Knut and Alice Wallenberg Foundation. This work was supported in part by NIH grants BICAN UM1 MH134812 and UM1 MH130981; BRAIN CONNECTS U01 NS132181 and UM1 NS132358; NIBIB R01 EB023281, R01 EB033773, R21 EB018907, R01 EB019956, and P41 EB030006; NICHD R00 HD101553; NIA R21 AG093037, R21 AG082082, R01 AG064027, R01 AG016495, and R01 AG070988; NIMH R01 MH123195, RF1 MH121885 and RF1 MH123195; NINDS U24 NS135561, R01 NS070963, R01 NS083534, R01 NS105820, and R25 NS125599; Blueprint for Neuroscience Research U01 MH093765; and Shared Instrumentation Grants S10 R023401, S10 R019307, and S10 R023043. The project benefited from computational hardware provided by the Massachusetts Life Sciences Center. 4.0.2 MH maintains a consulting relationship with Neuro42, Inc. BF is an advisor to DeepHealth. Their interests are reviewed and managed by Massachusetts General Hospital and Mass General Brigham in accordance with their conflict-of-interest policies. The authors have no other interests to declare that are relevant to the content of this article. References [1] S. M. Abulnaga, A. Hoopes, N. Dey, et al. (2025) MultiMorph: on-demand atlas construction. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 30906–30917. Cited by: §1. [2] M. Alzahrani et al. (2020) Audit feasibility for geometric distortion in magnetic resonance imaging for radiotherapy. Phys Imaging Radiat Oncol 15, p. 80–84. Cited by: §1. [3] V. Arsigny et al. (2006) A Log-Euclidean Framework for Statistics on Diffeomorphisms. In MICCAI 2006, p. 924–931. Cited by: §2.1. [4] J. Ashburner and K. J. Friston (2000) Voxel-based morphometry—the methods. NeuroImage 11 (6), p. 805–821. Cited by: §1. [5] J. Ashburner et al. (2000) Image Registration Using a Symmetric Prior-in Three Dimensions. Human Brain Mapping 9, p. 212–225. Cited by: §1. [6] B. B. Avants et al. (2008) Symmetric diffeomorphic image registration with cross-correlation: Evaluating automated labeling of elderly and neurodegenerative brain. Medical Image Analysis 12 (1), p. 26–41. Cited by: §1. [7] G. Balakrishnan et al. (2019) VoxelMorph: A Learning Framework for Deformable Medical Image Registration. IEEE Trans. on Medical Imaging 38 (8), p. 1788–1800. Cited by: §1. [8] J. Barnes et al. (2004) Differentiating AD from aging using semiautomated measurement of hippocampal atrophy rates. NeuroImage 23 (2), p. 574–581. Cited by: §1. [9] B. Billot et al. (2023) SynthSeg: Segmentation of brain MRI scans of any contrast and resolution without retraining. Medical Image Analysis 86, p. 102789. Cited by: §1, §2.1. [10] A. V. Dalca et al. (2019) Unsupervised learning of probabilistic diffeomorphic registration for images and surfaces. Medical Image Analysis 57, p. 226–236. Cited by: §1. [11] B. D. De Vos et al. (2019) A deep learning framework for unsupervised affine and deformable image registration. Medical Image Analysis 52, p. 128–143. Cited by: §1. [12] B. Fischl et al. (2002) Whole brain segmentation: automated labeling of neuroanatomical structures in the human brain. Neuron 33 (3), p. 341–355. Cited by: §1. [13] B. Fischl (2012) FreeSurfer. NeuroImage 62 (2), p. 774–781. Cited by: §3. [14] J. Fu et al. (2025) Decomposing the effect of normal aging and Alzheimer’s disease in brain morphological changes via learned aging templates. Sci Rep 15 (1), p. 11813. Cited by: §1. [15] J. Fu et al. (2025) Learning Accurate Rigid Registration for Longitudinal Brain MRI from Synthetic Data. In ISBI 2025, p. 1–5. Cited by: §1, §2.1. [16] J. Fu et al. (2025) Synthesizing individualized aging brains in health and disease with generative models and parallel transport. Medical Image Analysis 105, p. 103669. Cited by: §3.1, §3.2. [17] C. D. Good et al. (2001) A voxel-based morphometric study of ageing in 465 normal adult human brains. NeuroImage 14 (1), p. 21–36. Cited by: §1. [18] K. Gopinath et al. (2024) Registration by regression (RbR): A framework for interpretable and flexible atlas registration. In WBIR, p. 205–215. Cited by: §1. [19] D. Greve and B. Fischl (2024) The FreeSurfer Maintenance Dataset. OpenNeuro. Cited by: §3. [20] J. L. Gunter et al. (2009) Measurement of MRI scanner performance with the ADNI phantom. Med Phys 36 (6Part1), p. 2193–2205. Cited by: §1. [21] M. Hoffmann et al. (2022) SynthMorph: Learning Contrast-Invariant Registration Without Acquired Images. IEEE Trans. on Medical Imaging 41 (3), p. 543–558. Cited by: §1. [22] M. Hoffmann et al. (2024) Anatomy-aware and acquisition-agnostic joint registration with SynthMorph. Imaging Neurosci 2, p. imag–2–00197. Cited by: §1, §2.1. [23] M. Hoffmann (2025) Domain-Randomized Deep Learning for Neuroimage Analysis. IEEE Signal Process Mag 42 (4), p. 78–90. Cited by: §1. [24] A. Hoopes et al. (2022) SynthStrip: skull-stripping for any brain image. NeuroImage 260, p. 119474. Cited by: §1. [25] A. Hoopes et al. (2024) VoxelPrompt: A Vision Agent for End-to-End Medical Image Analysis. arXiv preprint arXiv:2410.08397. Cited by: §3. [26] J. E. Iglesias et al. (2023) SynthSR: A public AI tool to turn heterogeneous clinical brain scans into high-resolution T1-weighted images for 3D morphometry. Sci Adv 9 (5), p. eadd3607. Cited by: §1. [27] Information eXtraction from Images (IXI) Project (2015) IXI Dataset: nearly 600 MR images from normal, healthy subjects. Imperial College London, King’s College London, and University College London. Note: https://brain-development.org/ixi-dataset/ Cited by: §3. [28] C. R. Jack Jr. et al. (2008) The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods. J Magn Reson Imaging 27 (4), p. 685–691. Cited by: §3.1, §3.2, §3.2. [29] B. Jian et al. (2025) Disentangling progress in medical image registration: Beyond trend-driven architectures towards domain-specific strategies. arXiv preprint arXiv:2512.01913. Cited by: §3. [30] J. Jovicich et al. (2006) Reliability in multi-site structural MRI studies: effects of gradient non-linearity correction on phantom and human data. NeuroImage 30 (2), p. 436–443. Cited by: §1, §1, §3.2. [31] D. B. Keator et al. (2016) The Function Biomedical Informatics Research Network Data Repository. NeuroImage 124, p. 1074–1079. Cited by: Figure 1, Figure 1, §3.2, §3.2. [32] B. Khanal et al. (2016) A biophysical model of brain deformation to simulate and analyze longitudinal MRIs of patients with Alzheimer’s disease. NeuroImage 134, p. 35–52. Cited by: §3.1. [33] I. B. Malone et al. (2013) MIRIAD—Public release of a multiple time point Alzheimer’s MR imaging dataset. NeuroImage 70, p. 33–36. Cited by: §3.1. [34] M. I. Miller (2004) Computational anatomy: shape, growth, and atrophy comparison via diffeomorphisms. NeuroImage 23, p. S19–S33. Cited by: §1. [35] K. Nousiainen and T. Mäkelä (2020) Measuring geometric accuracy in magnetic resonance imaging with 3D-printed phantom and nonrigid image registration. MAGMA 33 (3), p. 401–410. Cited by: §1. [36] M. Reuter et al. (2010) Highly accurate inverse consistent registration: A robust approach. NeuroImage 53 (4), p. 1181–1196. Cited by: §1. [37] M. Reuter et al. (2012) Within-subject template estimation for unbiased longitudinal image analysis. NeuroImage 61 (4), p. 1402–1418. Cited by: §1, §3, §3. [38] S. L. Risacher et al. (2010) Longitudinal MRI atrophy biomarkers: relationship to conversion in the ADNI cohort. Neurobiol Aging 31 (8), p. 1401–1418. Cited by: §1, §3.1. [39] H. Takao et al. (2010) Computational analysis of cerebral cortex. Neuroradiol 52 (8), p. 691–698. Cited by: §1. [40] X. Tang et al. (2015) The diffeomorphometry of regional shape change rates and its relevance to cognitive deterioration in mild cognitive impairment and Alzheimer’s disease. Human Brain Mapping 36 (6), p. 2093–2117. Cited by: §3.1, §3.2. [41] J. P. Thirion (1998) Image matching as a diffusion process: an analogy with Maxwell’s demons.. Medical Image Analysis 2 (3), p. 243–260. Cited by: §1. [42] L. Tian et al. (2024) uniGradICON: A foundation model for medical image registration. In MICCAI, p. 749–760. Cited by: §1. [43] A. J. W. Van der Kouwe et al. (2008) Brain morphometry with multiecho MPRAGE. NeuroImage 40 (2), p. 559–569. Cited by: §1, §4. [44] P. A. Yushkevich et al. (2010) Bias in estimation of hippocampal atrophy using deformation-based morphometry arises from asymmetric global normalization: an illustration in ADNI 3 T MRI data.. NeuroImage 50 (2), p. 434–445. Cited by: §1.