Calibrating Scanner-Induced False Atrophy the Way Adaptive Optics Calibrates a Wavefront Sensor

Many measurements of how fast a brain is shrinking in Alzheimer's disease come from nonlinear registration. The method warps a patient's earlier MRI onto a later one and reads atrophy off the stretching and compression of the warp. The warp is estimated from image intensities, so anything that changes intensities without changing anatomy can show up as apparent tissue loss. Examples include a scanner upgrade, a change in field strength, a drifting coil sensitivity, or a different contrast between gray and white matter. High-power laser optics has an exact analogue and a mature fix. When a laser beam heats its own optics, the resulting "thermal lens" bends the wavefront. The sensor that measures that bending, a Hartmann sensor, can be fooled when uneven beam brightness shifts its spot centroids, and the resulting error looks like real wavefront tilt. Adaptive-optics engineers handle this by calibration. They deliberately "poke" the system with known perturbations, record the sensor's response, and build a response matrix that separates real distortion from instrument artifact.

We propose to import that calibration procedure into longitudinal MRI. The method applies controlled, anatomy-free intensity perturbations to scans and records the deformation each one falsely produces. Those false deformations become an explicit basis of "artifact modes," which are projected out of every longitudinal deformation field before any disease model sees it. We test whether this removes false atrophy in scan pairs with no biological change, while preserving the discrimination of Alzheimer's disease and the prediction of conversion from mild cognitive impairment in the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort.

Core Claim

Registration-derived atrophy contains a component caused by intensity changes. We claim this component can be characterized once, without disease labels, by perturbation calibration. Removing it should reduce spurious change more than intensity normalization alone does, and it should do so without eroding the disease signal. The optics supplies the procedure: a response matrix built by pushing known inputs through the full pipeline, not by fitting a model of the artifact. The neuroimaging supplies the question and the benchmark.

The motivating optical work is the Lawrence Livermore real-time correction system, described in Real-time wavefront correction system using a zonal deformable mirror and a Hartmann sensor. That system corrected thermally induced astigmatism and coma in kilowatt dye-laser chains. It took 5 waves peak-to-valley of aberration down to 0.09 waves RMS within six iterations, using a control law built on measured actuator-to-sensor responses. It also documented the "focal plane problem," in which intensity structure masquerades as tilt. We reuse both ideas: response-matrix calibration, and the explicit worry that the sensor reads intensity as geometry. We do not reuse the hardware, and the analogy is methodological only. No light is involved in the brain side of this proposal.

Where this sits in the Alzheimer's literature

Two ADNI lines of work motivate and constrain the proposal.

The first line models deformations directly. Disentangling normal aging from Alzheimer's disease in structural magnetic resonance images writes anatomy as normal aging, estimated from three years of healthy-control follow-up as a stationary velocity field, plus a residual "specific morphological process." It reports 91% sensitivity and 84% specificity for Alzheimer's disease versus controls. (A stationary velocity field is a smooth vector field whose integration gives an invertible warp.) That model treats the estimated velocity field as purely anatomical. We add a stage in front of it that removes the part of the field explained by intensity artifacts. If scanner effects are absorbed into the "specific" component, then its reported discrimination partly measures site and hardware.

The second line attacks bias statistically. GSplit LBI: Taming the Procedural Bias in Neuroimaging for Disease Prediction showed that spatial normalization in voxel-based morphometry introduces label-correlated "procedural bias." It then separated a predictive estimator (90.91% accuracy, 3T Alzheimer's versus controls) from a sparse lesion estimator. That work is the closest in spirit, and it differs from ours in three ways. It operates on cross-sectional gray-matter maps rather than longitudinal warps. It learns the separation through the disease labels, with a task-dependent gap parameter. And it does not say what physical cause the bias has. Our calibration never touches labels and names each removed mode by the perturbation that created it.

Accurate brain age prediction with lightweight deep neural networks is relevant as a cautionary example. Its brain-age gap correlated with age at r = −0.39 until an explicit post-hoc calibration brought this to 0.01. Uncorrected systematic error is ordinary in this field, and correcting it is accepted practice.

Research Program

The model is short enough to state directly. For a subject's scan pair, the observed velocity field $v_{\text{obs}}$ is taken to be

$$v_{\text{obs}} = v_{\text{bio}} + M a + \epsilon,$$

where $v_{\text{bio}}$ is true anatomical change, $M$ is a matrix whose columns are artifact modes (the velocity fields produced by unit intensity perturbations), $a$ is the unknown per-pair amplitude of each mode, and $\epsilon$ is registration noise.

We build $M$ by poking. For a calibration set of scan pairs we apply families of intensity-only perturbations to the second image and register the pair with the same pipeline used downstream, recording each resulting velocity field. The perturbation families are smooth multiplicative bias fields, global and regional changes in gray–white contrast, noise-level changes, and intensity transforms matched to field-strength differences. We call a perturbation "intensity-only" when it changes the image but not the anatomy. Averaging responses across subjects in a common template space, followed by an orthogonal decomposition, gives a small set of modes. This is exactly the response-matrix step of adaptive optics, with the registration algorithm playing the role of the Hartmann sensor.

Correction then estimates $a$ by least squares and subtracts $Ma$. The main risk is identifiability. If real atrophy, such as ventricular enlargement, lies partly in the span of $M$, projection removes biology. We handle this in three ways. First, we measure the overlap between the span of $M$ and the healthy-aging subspace from the Pennec-style model. Second, we restrict correction to modes whose principal angle with that subspace exceeds a threshold. Third, we report how much disease signal is lost as a function of that threshold rather than hiding the choice.

Before the corrected fields enter any disease model, three validation stages test them. The first uses scan pairs with no expected anatomical change. The second uses scans with synthetically injected atrophy of known size, where we check that the injected change is recovered after correction. The third uses the full ADNI disease models.

Baselines

The naive reference point is the uncorrected pipeline: diffeomorphic registration after standard skull stripping and intensity normalization, with atrophy read from the divergence of the velocity field. It is run on identical scan pairs and registration settings. It sets the floor for false change in no-change pairs.

The strongest intensity-side competitor is aggressive preprocessing harmonization: bias-field correction plus histogram matching of each follow-up scan to its baseline. The comparison is fair only if this baseline gets the same tuning effort we spend on calibration. We therefore tune its parameters on the same calibration subjects, selecting for minimum false change. Beating it means lower apparent change in no-change pairs at equal or better recovery of injected atrophy.

The strongest disease-side method is the aging-plus-specific-process model of Disentangling normal aging from Alzheimer's disease in structural magnetic resonance images, which reports 91% sensitivity and 84% specificity for Alzheimer's versus controls. We reimplement it with and without our correction stage, using identical subjects and folds. Beating it does not require higher raw accuracy. The target is equal discrimination with lower dependence on scanner, measured as the drop in accuracy when training and test sites or field strengths differ.

For label-driven bias removal we use GSplit LBI. It reports 90.91% accuracy on 3T Alzheimer's versus controls, and the file gives no numerical value for its feature-stability metric (multi-set Dice coefficient). We run it on the same subjects with its gap parameter tuned by nested cross-validation.

For prognosis we use the amyloid-PET-and-MRI shape feature of Prediction of Conversion from Mild Cognitive Impairment to Alzheimer's Disease Using Amyloid PET and Brain MR Imaging Data. It reports AUC 0.807 for 48-month conversion, a multivariate hazard ratio of 5.70, and the same AUC for plain PET amyloid uptake. Our MRI-only corrected atrophy is not expected to beat a PET feature. The honest target is to close part of the gap between uncorrected MRI atrophy and 0.807, and to add value beyond it in a joint Cox model.

The key ablation replaces the calibrated modes with an equal number of random smooth modes, or with principal components of control-group velocity fields. This isolates whether the gain comes from perturbation-derived artifact modes specifically, or from any low-rank projection.

Why not just…?

Why not just use better bias correction and intensity normalization? This alternative is a real threat and gets a direct test. Bias correction removes slowly varying multiplicative gain. It does not remove changes in gray–white contrast, which move the tissue boundaries the registration locks onto. That is the same failure the optics paper describes, where intensity structure shifts centroids even after gain correction. If the tuned harmonization baseline matches calibration on no-change pairs and injected atrophy, the calibration adds nothing and we will report that.

Why not just add scanner and site as covariates in the statistical model? Covariates remove average offsets between scanners but assume the artifact is the same scalar everywhere. Calibration produces spatially structured modes that differ by region. The cross-field-strength transfer test decides this: covariate adjustment should do well within the training sites and degrade on held-out hardware, while projected modes should transfer if they are truly properties of the registration.

Why not just use GSplit LBI? It separates bias only for a given prediction task, and it needs labels and a tuned gap parameter (GSplit LBI). It therefore cannot clean deformations for label-free uses such as trial endpoints or subtype discovery. It is also complementary rather than exclusive, so we test it both alone and stacked on our corrected fields.

Why not just take principal components of healthy-control deformations and call them nuisance? Those components are dominated by normal aging, which is real anatomy. Removing them would delete the signal the Pennec model deliberately keeps. This is exactly what the random/PCA ablation measures.

Proposed Experiments

The first decisive check is false change. On ADNI pairs where anatomy should not have changed, the corrected pipeline should cut mean absolute apparent hippocampal and ventricular change relative to both the naive and harmonized baselines. Such pairs include same-session repeat acquisitions and scans of the same subject at 1.5T and 3T, and their availability must be verified in the release we obtain. We report this residual as a fraction of the typical annual Alzheimer's atrophy rate, analogous to reporting residual waves RMS against a one-tenth-wave design target.

The second check is atrophy injection. We warp healthy-control scans with known regional atrophy fields and verify that recovery error does not rise after correction, sweeping the subspace-angle threshold to map the trade-off.

The third check is disease discrimination and transfer. On Alzheimer's versus controls and on converting versus stable MCI, we compare the Pennec-style specific process with and without correction. We do this within field strength and train-on-1.5T/test-on-3T, then compare conversion AUC and Cox hazard ratios with the shape-feature reference.

The proposal succeeds if false change falls clearly below the harmonized baseline, injected atrophy is preserved, and cross-hardware discrimination degrades less. It fails cleanly if tuned harmonization matches it, or if the random-mode ablation performs as well.