Mirror-Odd Trajectories: Separating Asymmetric Disease from Symmetric Aging in Longitudinal Brain MRI

Many diseases first show up as a loss of left–right balance. One hippocampus shrinks faster than the other, or a tumor grows on one side. Normal aging, scanner drift and whole-head intensity changes mostly affect both sides together. We propose a self-supervised representation for longitudinal scans that is split exactly into two parts. The first is a mirror-even part, which stays the same when the image is reflected across the midline. The second is a mirror-odd part, which flips sign when the image is reflected. Each part gets its own time-aware training signal. The core bet is that conversion risk lives mainly in how fast the odd part moves between visits, not in where it sits at any single visit. A patient's own built-in asymmetry is a fixed, anatomy-specific offset that confuses static asymmetry measures. When we track change over time, that fixed offset cancels.

The idea comes from a habit in condensed-matter physics. When classification is extended from simple on-site symmetries to spatial symmetries such as reflection and rotation, physicists sort quantities by how they transform under the symmetry. Some quantities only become meaningful once the spatial symmetry is taken seriously. Longitudinal medical representation learning has so far used only the simplest symmetry, translation in time. We bring in the next one, reflection in space, and ask whether it carries prognostic information that time-equivariance alone throws away.

Core Claim

The closest existing method is 3DTINC. It learns retinal OCT representations whose distance between two visits of the same patient is pushed to grow in proportion to the elapsed time. It uses no spatial symmetry. Its downstream prediction also uses a single scan, and it loses accuracy across scanners: ROCAUC falls from 0.792 on Zeiss Cirrus to 0.675 on Topcon.

Our claim is that the reflection group gives a free, exact decomposition of any encoder's output. Take an encoder $f$, a scan $x$ registered so the midsagittal plane is fixed, and its mirror image $Rx$. We define the even part $e(x) = \tfrac12(f(x)+f(Rx))$ and the odd part $o(x) = \tfrac12(f(x)-f(Rx))$. These are exactly invariant and sign-flipping under reflection by construction, so no penalty term is needed.

We train both parts with a 3DTINC-style time-proportional margin, using separate scale parameters. We then predict conversion from the odd part's velocity across a patient's first few visits, $\Delta o / \Delta t$. If this velocity adds prognostic information beyond the even part, the plain encoder, and a hand-crafted left–right volume difference, then spatial symmetry is a missing ingredient in longitudinal biomarker learning.

The physics literature plays two roles here. It supplies the method, and it gives an analogy that should not be oversold. The method we borrow is to sort representations into the irreducible pieces of a symmetry group, the habit behind the crystalline classifications in Nelson et al. and Lapierre et al.. Nelson et al. show that rotation symmetry protects a "returning Thouless pump" invariant that exists only because of the spatial symmetry, and that adding extra bands can erase it. The analogy we take, loosely, is this: a signal defined relative to a spatial symmetry can be real yet fragile, and in our case registration error plays the role of the trivializing bands. We do not claim any topological invariant in the strict sense. The note's gap analysis also floats ideas such as Hopf linkages of biomarker trajectories and time-vortex defects in disease. We judged those to be metaphors without a measurable object, and set them aside.

Research Program

Encoder and splitting. We build a 3D convolutional encoder over T1-weighted MRI that has been rigidly registered to a symmetric template, so that mirroring about the midline is a well-defined operation. The parity split is applied at the representation level. Training uses pairs of visits from the same patient. The even part's distance is pushed toward $\alpha_e \Delta t$ and the odd part's toward $\alpha_o \Delta t$, using the non-contrastive form from 3DTINC. We also add a regularizer that keeps the odd part's between-subject variance from collapsing.

Patient summary. For each patient we fit a per-patient linear slope to $e$ and to $o$ over the available visits. The slope tolerates irregular visit spacing and cancels each subject's fixed baseline asymmetry. Prediction feeds these slopes, alone and together with baseline plasma or clinical covariates, to a Random Survival Forest. This matches Chen et al., so the downstream learner is held fixed across all comparisons.

Main application: ADNI. The main test is ADNI mild cognitive impairment (MCI) conversion to Alzheimer's disease. Asymmetric medial temporal atrophy is plausible there, and the file documents both baseline-biomarker and longitudinal linear baselines.

Secondary test: Lumiere. The Lumiere glioma MRI series used by CRONOS is a strongly one-sided disease. There the odd part should dominate.

Controlled test: synthetic data. We modify Health MNIST from L-VAE. Shared drift plays the role of aging, and we inject a one-sided deformation that starts at a known time, so the true onset is known. This checks whether the odd-part velocity recovers onset time when the ground truth is available.

Baselines

Naive reference. The do-nothing reference is the strongest static result in the file. Chen et al. report AUROC 0.787 for 4-year MCI-to-AD conversion from a single baseline p-Tau217 measurement with a Random Survival Forest, versus 0.758 for amyloid PET, on 201 ADNI MCI subjects. We rerun this on the subset of those subjects who also have at least two MRIs. We keep the same forest settings and folds. Beating it means improving AUROC when MRI slopes are added to p-Tau217. MRI slopes alone matching it would be a secondary win.

Last-scan reference. The second simple reference uses only the last available scan's representation, mirroring the "last context image" baseline that CRONOS found hard to beat in imaging. For CRONOS, the file reports only an image-reconstruction score, about 92.8% SSIM, not a conversion score. No reported conversion result was found on Lacuna for this baseline on ADNI.

Strongest learned competitor. This is the same encoder trained with the 3DTINC loss but without the parity split, which is our main ablation. It has identical architecture, data, number of epochs and hyperparameter search budget, with 20 configurations each. The split is the only difference. No reported result was found on Lacuna for 3DTINC on ADNI; its reported figure is ROCAUC 0.792 on HARBOR OCT. A second ablation symmetrizes the input by random mirroring, which is the invariant choice and should erase the odd signal. A third keeps only the even part.

Linear longitudinal competitor. We use the globally aligned ADMM enrichment of Lu et al.. It reports MMSE RMSE 0.2928 versus 0.4188 for its baseline on ADNI, and we compare on that MMSE task as well.

Learned longitudinal competitor. We also run L-VAE with a patient-specific Gaussian-process prior, giving it the same visit data.

We count a result as beating a baseline only if the AUROC or concordance index (C-index) improvement exceeds the bootstrap 95% interval over 1,000 resamples on held-out folds.

Why not just…?

Why not just track a hand-crafted asymmetry index, such as left-minus-right hippocampal volume, over time? This is the real threat. If it matches our odd-part slope, the learned split adds nothing beyond a classical segmentation measure, and we will report that. We include it as a baseline, computed from standard segmentations and summarized with the same per-patient slope. The learned odd part could still win only if asymmetric change shows up outside the few structures people normally segment. The ablation where we restrict the encoder's input field to the hippocampus tests exactly this.

Why not just use mirror augmentation, as standard self-supervised learning does? Mirroring as augmentation teaches the model to ignore asymmetry. This is the spatial version of the failure 3DTINC documented for time: invariant methods such as BYOL and Barlow Twins fell to about 0.60 on external OCT data because they treat a patient's changing scans as the same view. Our symmetrized-input ablation measures this loss directly.

Why not just feed both hemispheres, or the whole image, to a stronger longitudinal model such as L-VAE or CRONOS? Such a model could in principle learn the split by itself. CRONOS, however, uses an unconstrained 3D U-Net that predicts how voxels move over time, and the file reports only image-reconstruction scores for it. The unsplit-encoder ablation tests whether hard-wiring the split helps at matched capacity. We expect the advantage to be largest with few subjects, where structure that is guaranteed by design saves data.

Proposed Experiments and Decisive Checks

Primary check. The parity-split slopes must improve 4-year conversion AUROC and C-index over the unsplit 3DTINC-style encoder. They must also improve over the hand-crafted asymmetry slope. Adding them to p-Tau217 must beat the 0.787 static reference on the shared subjects.

Mechanism check. The gain should come from the odd part. Odd-only should retain most of the benefit, and even-only should fall back toward the unsplit encoder. On Lumiere the odd part should dominate. On synthetic Health MNIST, the time at which the odd slope crosses a threshold should track the injected onset.

Robustness check. Global intensity and contrast shifts that affect both sides equally fall entirely into the even part, by construction. We therefore test cross-vendor transfer within ADNI, training on one scanner vendor and testing on the others. We predict that the odd part degrades less than the full representation, which addresses the scanner fragility seen in 3DTINC.

Failure mode. Registration error turns real anatomy into spurious odd signal. We perturb the midline estimate by 1–3 mm and degrees and report how fast the odd-part gain disappears. This is the practical analog of a fragile symmetry-defined invariant. If the gain does not survive realistic registration noise, the idea fails.