Back-In-Time Regularisation (BITR)

• Back-In-Time Regularisation (BITR) is a strategy that employs backward temporal transformations to regularize learning and inversion in time-evolving systems.
• In longitudinal MRI synthesis, BITR leverages bidirectional training by swapping baseline and follow-up scans to improve structural change prediction and reduce bias.
• In backward parabolic problems, BITR utilizes fractional time operators and frequency-band splitting to convert exponential ill-posedness into stable, algebraic growth.

Searching arXiv for the provided BITR-related papers and closely related work to ground the article. Back-In-Time Regularisation (BITR) denotes a class of temporally structured regularization strategies in which backward temporal transformations are used to stabilize or improve inference about time-evolving systems. In the arXiv literature considered here, the term appears in two technically distinct settings. In longitudinal brain MRI synthesis, BITR is introduced as a bidirectional training strategy for a 3D temporally aware diffusion model, where baseline and follow-up scans are randomly swapped so that the model learns both forward and backward mappings conditioned on signed elapsed time (Litrico et al., 3 Sep 2025). In inverse problems for backward parabolic equations, BITR is realized through fractional-in-time operators that replace exponentially ill-posed backward heat inversion by an “almost well posed” subdiffusion-based reconstruction framework, often combined with frequency-band splitting and discrepancy-principle calibration (Kaltenbacher et al., 2019).

1. Conceptual scope and motivating problems

In longitudinal medical image synthesis, the motivation for BITR arises from three limitations identified in prior work: interpolation-based methods such as DDM and DiffuseMorph generate intermediate states rather than predictive future scans; conditioning on absolute age in age-imbalanced datasets encourages learning age-specific appearance instead of the rate of change with respect to elapsed time; and 2D slice-based generators discard full 3D anatomical context, impairing coherent modeling of volumetric change such as ventricular expansion and hippocampal atrophy (Litrico et al., 3 Sep 2025).

In backward parabolic inverse problems, the motivation is different but structurally related. For the forward equation

$$u_t-\mathbb{L}u=0 \quad \text{in } \Omega\times(0,T),$$

with terminal data $u(x,T)=g(x)$, spectral inversion of the heat semigroup requires multiplication by $e^{\lambda_k T}$ mode by mode. Because this exponentially amplifies high-frequency noise, recovery of the initial state from terminal data is severely ill posed (Kaltenbacher et al., 2019).

These two usages share a common regularization intuition. Backward temporal structure is not introduced because retrodiction is itself the main endpoint. Rather, backward dynamics are used to constrain how time-conditioned changes are learned or inverted. In the MRI setting, backward prediction is explicitly stated to have limited clinical applications, yet it regularizes the model’s mapping between elapsed time and structural change. In the PDE setting, backward dynamics are the core inverse problem, and regularization is required to make inversion stable enough to be analyzable and computable.

2. BITR in temporally aware diffusion for longitudinal MRI

Within the 3D Temporally-Aware Diffusion Model (TADM-3D), BITR is defined on paired scans from the same subject. The baseline image at time $T_a$ is

$x_0 \equiv I_{T_a},$

the follow-up image at time $T_b=T_a+\Delta t$ is

$x_{\Delta t}\equiv I_{T_b},$

and the residual image is

$r_{\Delta t}=x_{\Delta t}-x_0.$

The model generates a residual $\hat r_{\Delta t}$ and reconstructs the predicted scan as

$\hat x_{\Delta t}=x_0+\hat r_{\Delta t}.$

Forward training therefore maps baseline to follow-up conditioned on $u(x,T)=g(x)$0, while backward training swaps scan roles and conditions on $u(x,T)=g(x)$1, equivalently learning $u(x,T)=g(x)$2 and $u(x,T)=g(x)$3 (Litrico et al., 3 Sep 2025).

BITR is implemented by randomly swapping the roles of baseline and follow-up with probability $u(x,T)=g(x)$4. If $u(x,T)=g(x)$5 denotes the sampled direction, then the model is trained on residuals $u(x,T)=g(x)$6 and conditions on signed elapsed time $u(x,T)=g(x)$7. This is not an auxiliary post hoc constraint; it is the training protocol itself. The paper states that this bidirectionality regularizes the mapping between elapsed time and structural change, reduces directional bias, and improves temporal awareness. It is also motivated as a means of reducing sensitivity to imbalances in age and time gaps, with external validation on NACC used to assess generalization.

Conditioning is explicitly patient specific. The baseline latent representation is

$u(x,T)=g(x)$8

where $u(x,T)=g(x)$9 is a 3D UNet-based encoder. Baseline age $e^{\lambda_k T}$0 at $e^{\lambda_k T}$1 and cognitive status $e^{\lambda_k T}$2 are also included. When BITR swaps direction, the scan chosen as the input for that iteration becomes the effective “baseline,” and its metadata are used. Time is encoded as elapsed interval $e^{\lambda_k T}$3, not absolute age, via positional encoding; at backward steps, the sign is flipped and $e^{\lambda_k T}$4 is encoded.

The generative backbone $e^{\lambda_k T}$5 is a 3D U-Net following an SRDiff-style configuration, operating on 3D MRI volumes. The encoder $e^{\lambda_k T}$6 is trained end-to-end with the DDPM, while the Brain-Age Estimator (BAE) $e^{\lambda_k T}$7 is pre-trained on the training set and then frozen. Reported optimization settings are AdamW with learning rate $e^{\lambda_k T}$8, weight decay $e^{\lambda_k T}$9, batch size $T_a$0, and a cosine learning rate scheduler. MRIs are linearly registered to MNI152 and skull-stripped with FSL. Training pairs span multiple follow-ups per baseline in OASIS-3 and include intervals up to approximately $T_a$1 years.

3. Objectives, inference, and empirical behavior in TADM-3D

The DDPM is trained by $T_a$2-prediction on noisy residuals. With conditioning

$T_a$3

the diffusion loss is

$T_a$4

The BAE imposes an age-consistency constraint through

$T_a$5

The total loss is

$T_a$6

Under BITR, the direction-averaged formulation makes explicit that the same diffusion and age-consistency losses are optimized for both signs of the time interval: $T_a$7 Crucially, BITR introduces no additional explicit loss term beyond bidirectional sampling. The paper also states that no explicit supervised image reconstruction loss and no cycle-consistency term are used (Litrico et al., 3 Sep 2025).

At inference time, generation is forward only. Given baseline $T_a$8, desired interval $T_a$9, metadata $x_0 \equiv I_{T_a},$0, and latent $x_0 \equiv I_{T_a},$1, the model denoises Gaussian noise conditioned on $x_0 \equiv I_{T_a},$2 to produce $x_0 \equiv I_{T_a},$3 and hence $x_0 \equiv I_{T_a},$4. Temporal realism is attributed to three mechanisms used jointly during training: explicit conditioning on $x_0 \equiv I_{T_a},$5, BAE guidance enforcing age-consistent outputs, and BITR bidirectional learning.

The reported evaluation uses image-based MSE and SSIM, as well as region volumetric MAE (%) in hippocampus, amygdala, lateral ventricles, thalamus, and CSF via SynthSeg 2.0, normalized to total brain volume. On the OASIS-3 internal test set, TADM-3D achieves MSE $x_0 \equiv I_{T_a},$6 and SSIM $x_0 \equiv I_{T_a},$7. The lowest volumetric errors reported for TADM-3D are hippocampus $x_0 \equiv I_{T_a},$8, amygdala $x_0 \equiv I_{T_a},$9, lateral ventricles $T_b=T_a+\Delta t$0, thalamus $T_b=T_a+\Delta t$1, and CSF $T_b=T_a+\Delta t$2. The paper states improvements versus BrLP, CounterSynth, and DaniNet across image and region metrics, including approximately $T_b=T_a+\Delta t$3 lower hippocampal error and approximately $T_b=T_a+\Delta t$4 lower thalamic error than baselines. On the external NACC test set, TADM-3D reports MSE $T_b=T_a+\Delta t$5 and SSIM $T_b=T_a+\Delta t$6, with strong performance across $T_b=T_a+\Delta t$7 regions, which is presented as evidence of robustness to out-of-distribution data (Litrico et al., 3 Sep 2025).

Setting	MSE	SSIM
TADM-3D	$T_b=T_a+\Delta t$8	$T_b=T_a+\Delta t$9
Without BITR	$x_{\Delta t}\equiv I_{T_b},$0	$x_{\Delta t}\equiv I_{T_b},$1
Without BAE	$x_{\Delta t}\equiv I_{T_b},$2	$x_{\Delta t}\equiv I_{T_b},$3
Without age-gap conditioning	$x_{\Delta t}\equiv I_{T_b},$4	$x_{\Delta t}\equiv I_{T_b},$5
TADM-2.5D	$x_{\Delta t}\equiv I_{T_b},$6	$x_{\Delta t}\equiv I_{T_b},$7

The ablations identify BITR and BAE as complementary. Without BITR, region errors worsen, including hippocampus $x_{\Delta t}\equiv I_{T_b},$8 versus $x_{\Delta t}\equiv I_{T_b},$9 and thalamus $r_{\Delta t}=x_{\Delta t}-x_0.$0 versus $r_{\Delta t}=x_{\Delta t}-x_0.$1. Without BAE, volumetric errors also increase across regions. The larger degradation observed when removing age-gap conditioning is used to support conditioning on $r_{\Delta t}=x_{\Delta t}-x_0.$2 rather than absolute age. The 2.5D ablation is used to support the necessity of native 3D context. Qualitative comparisons are described as showing more realistic ventricular expansion and lower error heatmaps than BrLP, CounterSynth, and DaniNet.

4. Fractional-operator BITR for backward parabolic equations

In backward parabolic inversion, BITR regularizes the recovery of an initial state from terminal data. Let $r_{\Delta t}=x_{\Delta t}-x_0.$3 be bounded with $r_{\Delta t}=x_{\Delta t}-x_0.$4 boundary, and let $r_{\Delta t}=x_{\Delta t}-x_0.$5 be a uniformly elliptic second-order operator with homogeneous Dirichlet boundary conditions. With $r_{\Delta t}=x_{\Delta t}-x_0.$6 an $r_{\Delta t}=x_{\Delta t}-x_0.$7-orthonormal eigenbasis of $r_{\Delta t}=x_{\Delta t}-x_0.$8, $r_{\Delta t}=x_{\Delta t}-x_0.$9, the forward heat solution is

$\hat r_{\Delta t}$0

Since

$\hat r_{\Delta t}$1

naive backward reconstruction requires

$\hat r_{\Delta t}$2

which exponentially amplifies measurement noise in high-frequency modes (Kaltenbacher et al., 2019).

The paper replaces the classical parabolic operator by a fractional-in-time subdiffusion model using the Caputo derivative

$\hat r_{\Delta t}$3

written in the paper as $\hat r_{\Delta t}$4. The regularized dynamics are

$\hat r_{\Delta t}$5

Mode by mode, the solution is

$\hat r_{\Delta t}$6

where $\hat r_{\Delta t}$7 is the Mittag–Leffler function. Terminal coefficients $\hat r_{\Delta t}$8 therefore satisfy

$\hat r_{\Delta t}$9

and fractional inversion becomes

$\hat x_{\Delta t}=x_0+\hat r_{\Delta t}.$0

The crucial regularization effect is spectral. Heat inversion multiplies by $\hat x_{\Delta t}=x_0+\hat r_{\Delta t}.$1, whereas fractional inversion multiplies by the reciprocal of a Mittag–Leffler factor. The paper uses the bounds

$\hat x_{\Delta t}=x_0+\hat r_{\Delta t}.$2

to show that the growth is algebraic rather than exponential in frequency. This is the basis for the characterization of the backward subdiffusion problem as “almost well posed.” The paper further states that subdiffusion loses only two spatial derivatives and provides the regularity estimate

$\hat x_{\Delta t}=x_0+\hat r_{\Delta t}.$3

Fractional BITR is presented as analogous in spirit to classical quasireversibility, but with a nonlocal time operator. The comparison emphasized in the paper is that replacing the exponential heat factor by a Mittag–Leffler factor yields milder ill-conditioning and better high-frequency control.

5. Frequency-band calibration, convergence theory, and numerical reconstruction

The paper does not advocate a single uniform fractional order across the full spectrum. Instead, it develops a frequency-band splitting strategy motivated by the observation that fractional inversion can over-amplify low frequencies relative to exact parabolic inversion while being advantageous at high frequencies. In a three-band formulation, low frequencies $\hat x_{\Delta t}=x_0+\hat r_{\Delta t}.$4 are inverted by the exact parabolic factor, middle frequencies $\hat x_{\Delta t}=x_0+\hat r_{\Delta t}.$5 are inverted by a fractional operator of order $\hat x_{\Delta t}=x_0+\hat r_{\Delta t}.$6, and high frequencies $\hat x_{\Delta t}=x_0+\hat r_{\Delta t}.$7 are inverted by order $\hat x_{\Delta t}=x_0+\hat r_{\Delta t}.$8, often with $\hat x_{\Delta t}=x_0+\hat r_{\Delta t}.$9. The resulting modal reconstruction is

$u(x,T)=g(x)$00

Here $u(x,T)=g(x)$01 are coefficients of noisy terminal data $u(x,T)=g(x)$02, while $u(x,T)=g(x)$03 are coefficients of a smoothed version of the data (Kaltenbacher et al., 2019).

Smoothing is performed with a Landweber-type iteration

$u(x,T)=g(x)$04

with $u(x,T)=g(x)$05 and $u(x,T)=g(x)$06 chosen so that $u(x,T)=g(x)$07 for $u(x,T)=g(x)$08. An a priori choice

$u(x,T)=g(x)$09

yields

$u(x,T)=g(x)$10

The role of smoothing is explicit: because fractional inversion grows like $u(x,T)=g(x)$11 at large frequencies, the final data used in the fractional bands should lie in $u(x,T)=g(x)$12.

The discrepancy principle is used to select both spectral cutoffs and fractional orders. For the low-frequency cutoff,

$u(x,T)=g(x)$13

where

$u(x,T)=g(x)$14

For a fractional order $u(x,T)=g(x)$15, the paper chooses $u(x,T)=g(x)$16 so that

$u(x,T)=g(x)$17

with $u(x,T)=g(x)$18. Existence of such an $u(x,T)=g(x)$19 is proved under a signal-to-noise assumption and continuity of the residual map.

The convergence analysis separates a noise term and an approximation term: $u(x,T)=g(x)$20 where

$u(x,T)=g(x)$21

Under an a priori choice with $u(x,T)=g(x)$22 and $u(x,T)=g(x)$23, Theorem 4.1 gives strong $u(x,T)=g(x)$24 convergence. With $u(x,T)=g(x)$25 chosen a posteriori by the discrepancy principle, Theorem 4.2 gives weak convergence in $u(x,T)=g(x)$26. Under stronger source conditions and the choice $u(x,T)=g(x)$27, Theorem 4.3 derives logarithmic rates,

$u(x,T)=g(x)$28

and an analogous noise-free rate in terms of $u(x,T)=g(x)$29. For the split-frequency method, Theorem 5.1 establishes weak convergence of $u(x,T)=g(x)$30 to $u(x,T)=g(x)$31.

The numerical experiments are one-dimensional with $u(x,T)=g(x)$32, Dirichlet boundary conditions, final time $u(x,T)=g(x)$33, and noise levels $u(x,T)=g(x)$34 and $u(x,T)=g(x)$35. The compared strategies include pure truncated SVD, single split, double split, and triple split reconstructions. In Example 1, a double split with $u(x,T)=g(x)$36, $u(x,T)=g(x)$37, $u(x,T)=g(x)$38, and $u(x,T)=g(x)$39 improves resolution near a derivative jump relative to a single split with $u(x,T)=g(x)$40 and $u(x,T)=g(x)$41. In Example 2, triple split reconstruction is reported to outperform truncated SVD and single split by capturing more mid- and high-frequency structure. The advantage is stated to increase as $u(x,T)=g(x)$42 decreases or $u(x,T)=g(x)$43 decreases.

6. Related methods, misconceptions, and limitations

In longitudinal MRI synthesis, BITR is contrasted with interpolation-based diffusion methods such as DDM and DiffuseMorph, which generate intermediate states along deformation paths rather than future scans at arbitrary $u(x,T)=g(x)$44. It is also contrasted with sequence-conditioned diffusion such as SADM, which requires longitudinal sequences at inference, whereas TADM-3D requires only a baseline scan and desired $u(x,T)=g(x)$45. Relative to 2D slice generators, the argument is that bidirectional temporal consistency is best exploited in a native 3D backbone preserving global volumetric structure. A frequent misconception would be to equate BITR in TADM-3D with a cycle-consistency method; the paper explicitly states the opposite. BITR there is stochastic direction swapping inside the same diffusion objective, with no extra cycle loss and no explicit supervised reconstruction term. Another misconception would be to treat backward generation as the intended clinical product; the paper states that clinical inference is forward only, and backward generation is used only during training as regularization (Litrico et al., 3 Sep 2025).

In backward parabolic inversion, BITR is situated against classical quasireversibility, truncated SVD, Tikhonov regularization, and space-fractional pseudoparabolic alternatives. The paper’s conclusion favors time-fractional inversion, especially in conjunction with band splitting, because it replaces exponential spectral amplification by algebraic growth and empirically provides better recovery of mid- and high-frequency structure. Tikhonov is specifically not used for smoothing because of saturation; Landweber iteration is preferred for placing the data in $u(x,T)=g(x)$46 or $u(x,T)=g(x)$47. Space-fractional pseudoparabolic regularization and combined space-time fractional variants are analyzed, but sensitivity to $u(x,T)=g(x)$48 is identified as a limitation. A corresponding misconception would be to interpret fractional BITR as exact inversion of the original heat equation; it is instead a surrogate inverse framework with its own dynamics, stability theory, and calibration rules (Kaltenbacher et al., 2019).

The main limitations are domain specific. For TADM-3D, long-term predictions at large $u(x,T)=g(x)$49 are challenging, BAE inaccuracies can modulate the age-consistency loss, conditioning is limited to age and cognitive status, and distribution shift remains nontrivial even though external NACC validation shows robustness. For fractional BITR, effective deployment requires a reliable estimate of the noise level $u(x,T)=g(x)$50, accurate knowledge of $u(x,T)=g(x)$51 and boundary conditions, and acceptance that very large $u(x,T)=g(x)$52 or large $u(x,T)=g(x)$53 may force severe truncation and loss of fine features. A plausible common implication is that BITR is most effective when backward temporal structure is used as a stabilizing inductive bias and when the remaining sources of mismatch—dataset imbalance in one case, operator and noise uncertainty in the other—are explicitly controlled.