Time-Conditioned AR Contrast Enhancement

• T-CACE is a modeling principle for MRI contrast enhancement that decomposes the process into sequential, time-conditioned autoregressive steps to maintain anatomical integrity.
• It leverages diverse architectures such as ConvLSTM, masked self-attention, and neural cellular automata to achieve coherent phase-wise enhancement and robust multi-task performance.
• Empirical results across kidney, brain, liver, and breast MRI demonstrate improved synthesis, segmentation, and classification metrics, validating its clinical potential.

Time-Conditioned Autoregressive Contrast Enhancement (T-CACE) denotes a class of formulations in which MRI contrast enhancement is modeled as an ordered, time-conditioned generation process rather than as a single static image-to-image mapping. Across the literature, the central idea is to decompose synthesis into temporally or phase-ordered subproblems and to propagate latent, token, or state representations forward through those steps so that enhancement evolves coherently while anatomical content remains stable. The term is used explicitly for a spatio-temporal DCE-MRI style-transfer framework in which contrast dynamics are modeled as a time-conditioned autoregressive process in latent space (Tattersall et al., 2023), and it is also the name of a later multi-task liver MRI framework that jointly synthesizes multi-phase contrast-enhanced MRI, segments lesions, and performs tumor classification (Xiao et al., 13 Aug 2025). Closely related work recasts discrete contrast-agent dose levels as synthetic time steps (Gui et al., 2024), implements enhancement dynamics through iterative neural cellular automata (Lang et al., 23 Jun 2025), or motivates autoregressive extensions from a non-autoregressive kinetics world model with explicit spatiotemporal consistency learning (Kong et al., 22 Feb 2026).

1. Conceptual scope and lineages

T-CACE is best understood as a modeling principle for contrast dynamics. The “time” variable may correspond to real DCE acquisition frames, synthetic enhancement steps, ordered clinical phases such as arterial, portal venous, and delayed imaging, or a continuous elapsed-time variable after contrast administration. The “autoregressive” component means that the representation at step $t$ depends on prior steps, typically through a recurrent state, a causal attention mask, or an iterative state update. The “contrast enhancement” component refers not merely to intensity amplification, but to the spatially localized and physiologically ordered appearance of enhancement across tissues.

Paper	Realization of T-CACE	Imaging setting
(Tattersall et al., 2023)	ConvLSTM over disentangled content latents with adaptive convolution decoding	Kidney 2D DCE-MRI; prostate 3D DCE-MRI
(Gui et al., 2024)	Masked self-attention over dose/time-variant tokens with Swin UNETR tokenizer	Brain tumor virtual contrast-enhanced MRI
(Xiao et al., 13 Aug 2025)	Conditional Token Encoding, Dynamic Time-Aware Attention Mask, Temporal Classification Consistency	Liver multi-phase MRI synthesis, segmentation, diagnosis
(Lang et al., 23 Jun 2025)	ODE-like iterative neural cellular automaton rollout	Breast DCE-MRI
(Kong et al., 22 Feb 2026)	Time-conditioned latent diffusion with Latent Alignment Learning and Latent Difference Learning	Abdominal and breast contrast-enhancement kinetics

The earliest formulation in this set operationalizes T-CACE through disentangled content and style latents, with a bidirectional ConvLSTM propagating content information across adjacent time points and adaptive convolutions injecting localized enhancement behavior during decoding (Tattersall et al., 2023). A distinct transformer-based line reinterprets progressive contrast-agent dosage as a discrete enhancement trajectory and then maps those dose steps to a time-conditioned autoregressive token model (Gui et al., 2024). In liver MRI, T-CACE is expanded into a unified multi-task framework in which phase-conditioned synthesis is coupled to lesion segmentation and classification, and temporal conditioning is enforced both in attention and in a classification-consistency constraint (Xiao et al., 13 Aug 2025). Related breast and kinetics-world-model work broadens the concept to irregularly sampled temporal sequences and continuous-time generation (Lang et al., 23 Jun 2025, Kong et al., 22 Feb 2026).

This suggests that T-CACE is not a single architecture. Rather, it is a family of temporally structured enhancement models whose shared concern is the preservation of anatomy under evolving contrast dynamics.

2. Probabilistic and dynamical formulations

A common formalization writes the enhancement trajectory as a conditional sequential factorization. In the phase- or time-conditioned transformer setting, the image sequence can be expressed as

$$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$

or, in token space,

$$p(z_{\mathrm{var}}^{(1:T)}\mid z_{\mathrm{inv}},x,c)=\prod_{t=1}^{T} p\big(z_{\mathrm{var}}^{(t)}\mid z_{\mathrm{var}}^{(<t)},z_{\mathrm{inv}},x,c\big),$$

where $z_{\mathrm{inv}}$ is fixed across steps and $z_{\mathrm{var}}$ carries the evolving enhancement-dependent information (Gui et al., 2024). In the explicit liver T-CACE factorization, the three clinical phases are generated in order:

$$p(y_{\mathrm{Art}},y_{\mathrm{PV}},y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}}) = p(y_{\mathrm{Art}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},t_{\mathrm{Art}}) \cdot p(y_{\mathrm{PV}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},t_{\mathrm{PV}}) \cdot p(y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},y_{\mathrm{PV}},t_{\mathrm{Delay}}),$$

making phase ordering part of the probabilistic structure (Xiao et al., 13 Aug 2025).

A latent recurrent variant models temporal content dynamics as

$$p(z_{c,t}\mid z_{c,<t},u_t)=p(z_{c,t}\mid z_{c,t-1},u_t),$$

where $u_t$ is a time-conditioning signal. In the reported implementation, $u_t$ is implicit in ordered ConvLSTM recurrence rather than an explicit embedding, and a deterministic decoder $G$ reconstructs the frame through

$$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$0

Teacher forcing aligns predicted latent states $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$1 to encoder-produced $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$2 through an autoregressive temporal loss (Tattersall et al., 2023).

The neural cellular automaton realization takes a dynamical-systems form. The global state $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$3 is advanced by an Euler-like update,

$$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$4

with the visible channel initialized from the pre-contrast image and the hidden channels serving as internal memory. Here time-conditioning is injected directly through the step size $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$5, fixed at $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$6 seconds in the reported experiments (Lang et al., 23 Jun 2025). The resulting process is deterministic autoregression over states rather than over token sequences.

A crucial boundary case is the kinetics world-model approach, which conditions on absolute time and a non-contrast image but does not implement an explicit transition of the form $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$7. Its mapping is instead

$$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$8

with temporal smoothness imposed by latent regularization rather than by an autoregressive transition operator (Kong et al., 22 Feb 2026). This distinction matters because T-CACE is sometimes used as an interpretive framework for such models rather than as a literal description of the trained generator.

3. Architectural realizations

The recurrent autoencoder realization separates anatomy from enhancement-dependent appearance. Shared-weight encoders produce content and style codes,

$$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$9

and a bidirectional ConvLSTM evolves the content bottleneck over five-frame windows. The recurrent cell uses convolutional gates,

$$p(z_{\mathrm{var}}^{(1:T)}\mid z_{\mathrm{inv}},x,c)=\prod_{t=1}^{T} p\big(z_{\mathrm{var}}^{(t)}\mid z_{\mathrm{var}}^{(<t)},z_{\mathrm{inv}},x,c\big),$$0

$$p(z_{\mathrm{var}}^{(1:T)}\mid z_{\mathrm{inv}},x,c)=\prod_{t=1}^{T} p\big(z_{\mathrm{var}}^{(t)}\mid z_{\mathrm{var}}^{(<t)},z_{\mathrm{inv}},x,c\big),$$1

$$p(z_{\mathrm{var}}^{(1:T)}\mid z_{\mathrm{inv}},x,c)=\prod_{t=1}^{T} p\big(z_{\mathrm{var}}^{(t)}\mid z_{\mathrm{var}}^{(<t)},z_{\mathrm{inv}},x,c\big),$$2

with $$p(z_{\mathrm{var}}^{(1:T)}\mid z_{\mathrm{inv}},x,c)=\prod_{t=1}^{T} p\big(z_{\mathrm{var}}^{(t)}\mid z_{\mathrm{var}}^{(<t)},z_{\mathrm{inv}},x,c\big),$$3 and $$p(z_{\mathrm{var}}^{(1:T)}\mid z_{\mathrm{inv}},x,c)=\prod_{t=1}^{T} p\big(z_{\mathrm{var}}^{(t)}\mid z_{\mathrm{var}}^{(<t)},z_{\mathrm{inv}},x,c\big),$$4 or a linear projection thereof. This recurrence is paired with adaptive convolution, where style codes generate depth-wise kernels, point-wise kernels, and biases that modulate content features spatially and channel-wise. The stated motivation is that global style transfer methods such as AdaIN smear contrast enhancement across all tissues, whereas adaptive convolution respects the localized nature of enhancement (Tattersall et al., 2023).

The transformer realization decomposes tokens into invariant and variant components. A Swin UNETR-based tokenizer produces $$p(z_{\mathrm{var}}^{(1:T)}\mid z_{\mathrm{inv}},x,c)=\prod_{t=1}^{T} p\big(z_{\mathrm{var}}^{(t)}\mid z_{\mathrm{var}}^{(<t)},z_{\mathrm{inv}},x,c\big),$$5 and $$p(z_{\mathrm{var}}^{(1:T)}\mid z_{\mathrm{inv}},x,c)=\prod_{t=1}^{T} p\big(z_{\mathrm{var}}^{(t)}\mid z_{\mathrm{var}}^{(<t)},z_{\mathrm{inv}},x,c\big),$$6, and masked self-attention is then applied across the concatenated token sequence of all dose or time steps. The staircase causal mask permits a step-$$p(z_{\mathrm{var}}^{(1:T)}\mid z_{\mathrm{inv}},x,c)=\prod_{t=1}^{T} p\big(z_{\mathrm{var}}^{(t)}\mid z_{\mathrm{var}}^{(<t)},z_{\mathrm{inv}},x,c\big),$$7 token to attend to all tokens from steps $$p(z_{\mathrm{var}}^{(1:T)}\mid z_{\mathrm{inv}},x,c)=\prod_{t=1}^{T} p\big(z_{\mathrm{var}}^{(t)}\mid z_{\mathrm{var}}^{(<t)},z_{\mathrm{inv}},x,c\big),$$8 while preventing access to future steps. Intermediate steps are decoded and re-encoded through an auxiliary encoder $$p(z_{\mathrm{var}}^{(1:T)}\mid z_{\mathrm{inv}},x,c)=\prod_{t=1}^{T} p\big(z_{\mathrm{var}}^{(t)}\mid z_{\mathrm{var}}^{(<t)},z_{\mathrm{inv}},x,c\big),$$9, so that updated lower-dose or earlier-time tokens are used as context for later predictions. In the dose-conditioned form, the reported steps are low dose, high dose, and standard dose; in the T-CACE reinterpretation, these become early, middle, and later enhancement steps (Gui et al., 2024).

The liver multi-task formulation extends this transformer pattern by injecting explicit temporal priors into the token itself. Conditional Token Encoding is defined as

$z_{\mathrm{inv}}$0

$z_{\mathrm{inv}}$1

At each phase, masked multi-head self-attention with Swin window position encoding consumes the current conditional token together with previous synthesized phase tokens. The Dynamic Time-Aware Attention Mask multiplies attention by a Gaussian decay,

$z_{\mathrm{inv}}$2

with $z_{\mathrm{inv}}$3, so that nearby phases exert stronger influence than temporally distant ones. Decoding yields both a synthesized phase image and a segmentation mask, and majority voting across phase-wise segmentations produces the final lesion mask (Xiao et al., 13 Aug 2025).

The neural cellular automaton realization dispenses with global attention and instead relies on repeated local message passing. The state has $z_{\mathrm{inv}}$4 channels, comprising one visible channel and $z_{\mathrm{inv}}$5 hidden channels. Perception at cell $z_{\mathrm{inv}}$6 concatenates the responses of two learnable $z_{\mathrm{inv}}$7 kernels and the identity pathway:

$z_{\mathrm{inv}}$8

A two-layer MLP with hidden size $z_{\mathrm{inv}}$9 predicts the state derivative, which is multiplied by a stochastic binary fire mask $z_{\mathrm{var}}$0 before the Euler-like update is applied. Losses are computed only when the cumulative rollout time equals an available acquisition time, which directly addresses non-uniform sampling (Lang et al., 23 Jun 2025).

The kinetics world-model line uses a latent diffusion U-Net with ControlNet-style conditioning from the non-contrast image and a CLIP time encoder applied to absolute elapsed time represented as text in the form “HH:MM:SS”. Its SpatioTemporal Consistency Learning combines Latent Alignment Learning, which constructs a patient-specific latent covariance template, with Latent Difference Learning, which inserts interpolated time points and penalizes second-order latent curvature on the densified timeline. The paper explicitly states that this model is non-autoregressive, but it also provides a concrete latent-autoregressive T-CACE extension based on $z_{\mathrm{var}}$1 and reuse of the same regularizers (Kong et al., 22 Feb 2026).

4. Optimization objectives and temporal constraints

T-CACE methods are distinguished not only by their transition operators but also by how they prevent anatomy drift, enhancement leakage, temporal flicker, and physiologically implausible trajectories.

In the recurrent style-transfer formulation, the total objective is

$z_{\mathrm{var}}$2

Here $z_{\mathrm{var}}$3 is frame reconstruction, $z_{\mathrm{var}}$4 is encode-decode-re-encode consistency for content and style, $z_{\mathrm{var}}$5 is least-squares GAN realism, $z_{\mathrm{var}}$6 is perceptual feature preservation using VGG-16 or MedicalNet ResNet-18, $z_{\mathrm{var}}$7 is Gram-matrix style reconstruction, $z_{\mathrm{var}}$8 aligns ConvLSTM predictions with encoder latents through teacher forcing, and $z_{\mathrm{var}}$9 penalizes latent changes while relaxing smoothness where style changes rapidly (Tattersall et al., 2023).

In the transformer tokenizer setting, optimization is separated into tokenizer pretraining and autoregression training. The tokenizer uses an autoencoding loss and a prediction loss, each combining an $$p(y_{\mathrm{Art}},y_{\mathrm{PV}},y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}}) = p(y_{\mathrm{Art}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},t_{\mathrm{Art}}) \cdot p(y_{\mathrm{PV}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},t_{\mathrm{PV}}) \cdot p(y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},y_{\mathrm{PV}},t_{\mathrm{Delay}}),$$0 term with an adversarial term using a 3D discriminator adapted from ResViT. The autoregressive stage supervises every intermediate output with a mixture of $$p(y_{\mathrm{Art}},y_{\mathrm{PV}},y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}}) = p(y_{\mathrm{Art}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},t_{\mathrm{Art}}) \cdot p(y_{\mathrm{PV}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},t_{\mathrm{PV}}) \cdot p(y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},y_{\mathrm{PV}},t_{\mathrm{Delay}}),$$1 regression and MAE,

$$p(y_{\mathrm{Art}},y_{\mathrm{PV}},y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}}) = p(y_{\mathrm{Art}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},t_{\mathrm{Art}}) \cdot p(y_{\mathrm{PV}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},t_{\mathrm{PV}}) \cdot p(y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},y_{\mathrm{PV}},t_{\mathrm{Delay}}),$$2

and teacher forcing is implemented through decode-then-re-encode feedback, so that the updated intermediate tokens correct the variant pathway before the next step (Gui et al., 2024).

The liver framework couples four task losses:

$$p(y_{\mathrm{Art}},y_{\mathrm{PV}},y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}}) = p(y_{\mathrm{Art}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},t_{\mathrm{Art}}) \cdot p(y_{\mathrm{PV}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},t_{\mathrm{PV}}) \cdot p(y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},y_{\mathrm{PV}},t_{\mathrm{Delay}}),$$3

The synthesis term is an $$p(y_{\mathrm{Art}},y_{\mathrm{PV}},y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}}) = p(y_{\mathrm{Art}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},t_{\mathrm{Art}}) \cdot p(y_{\mathrm{PV}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},t_{\mathrm{PV}}) \cdot p(y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},y_{\mathrm{PV}},t_{\mathrm{Delay}}),$$4 loss over the three generated phases. The segmentation term combines Dice and cross-entropy. The classification term is cross-entropy on the lesion label. The Temporal Classification Consistency term compares per-phase classification predictions to a thresholded scalar signal predicted from latent non-contrast features and time:

$$p(y_{\mathrm{Art}},y_{\mathrm{PV}},y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}}) = p(y_{\mathrm{Art}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},t_{\mathrm{Art}}) \cdot p(y_{\mathrm{PV}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},t_{\mathrm{PV}}) \cdot p(y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},y_{\mathrm{PV}},t_{\mathrm{Delay}}),$$5

with $$p(y_{\mathrm{Art}},y_{\mathrm{PV}},y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}}) = p(y_{\mathrm{Art}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},t_{\mathrm{Art}}) \cdot p(y_{\mathrm{PV}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},t_{\mathrm{PV}}) \cdot p(y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},y_{\mathrm{PV}},t_{\mathrm{Delay}}),$$6. All loss weights are set to $$p(y_{\mathrm{Art}},y_{\mathrm{PV}},y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}}) = p(y_{\mathrm{Art}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},t_{\mathrm{Art}}) \cdot p(y_{\mathrm{PV}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},t_{\mathrm{PV}}) \cdot p(y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},y_{\mathrm{PV}},t_{\mathrm{Delay}}),$$7 (Xiao et al., 13 Aug 2025).

The cellular-automaton formulation is substantially lighter. It uses end-to-end free-run rollout without teacher forcing or scheduled sampling, and the reported training loss is simply the MSE between generated visible states and the available post-contrast targets at those rollout indices that correspond to actual acquisition times (Lang et al., 23 Jun 2025). By contrast, the kinetics world model optimizes a diffusion noise-prediction loss together with spatial and temporal regularizers in a two-stage schedule:

$$p(y_{\mathrm{Art}},y_{\mathrm{PV}},y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}}) = p(y_{\mathrm{Art}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},t_{\mathrm{Art}}) \cdot p(y_{\mathrm{PV}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},t_{\mathrm{PV}}) \cdot p(y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},y_{\mathrm{PV}},t_{\mathrm{Delay}}),$$8

Its spatial term aligns each time point to a patient-specific latent covariance template, while its temporal term penalizes second-order latent differences over a densified time grid (Kong et al., 22 Feb 2026).

Taken together, these objectives indicate that T-CACE is defined as much by its regularization strategy as by its predictor. Temporal coherence is enforced through teacher forcing, causal masking, Gaussian-decayed phase interactions, adaptive loss evaluation at irregular times, or latent smoothness penalties, depending on the sampling regime and the intended clinical task.

5. Empirical behavior across organs, datasets, and tasks

Reported evaluations span multiple organs and substantially different problem settings, so direct numerical comparison across papers is not straightforward. This suggests that T-CACE should be judged within each acquisition protocol and downstream task rather than through a single aggregate benchmark.

In dynamic kidney and prostate MRI style transfer, the spatio-temporal recurrent framework is evaluated on a Kidney 2D DCE-MRI dataset of 13 patients and 375 images and a Prostate 3D DCE-MRI dataset of 20 patients and 40 volumes, each with patient-level 80:20 splitting and five-fold cross-validation. Baselines are MUNIT, CycleGAN, and StyleGAN3, with DRIT++ excluded for computational reasons with 3D data. The method consistently outperforms the baselines in PSNR, SSIM, MS-SSIM, and both content/style CW-SSIM. Representative averages include Kidney Non-CE$$p(y_{\mathrm{Art}},y_{\mathrm{PV}},y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}}) = p(y_{\mathrm{Art}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},t_{\mathrm{Art}}) \cdot p(y_{\mathrm{PV}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},t_{\mathrm{PV}}) \cdot p(y_{\mathrm{Delay}}\mid x_{\mathrm{NC}},x_{\mathrm{TM}},y_{\mathrm{Art}},y_{\mathrm{PV}},t_{\mathrm{Delay}}),$$9CE PSNR $$p(z_{c,t}\mid z_{c,<t},u_t)=p(z_{c,t}\mid z_{c,t-1},u_t),$$0 for the proposed method versus $$p(z_{c,t}\mid z_{c,<t},u_t)=p(z_{c,t}\mid z_{c,t-1},u_t),$$1 for MUNIT, $$p(z_{c,t}\mid z_{c,<t},u_t)=p(z_{c,t}\mid z_{c,t-1},u_t),$$2 for CycleGAN, and $$p(z_{c,t}\mid z_{c,<t},u_t)=p(z_{c,t}\mid z_{c,t-1},u_t),$$3 for StyleGAN3; Kidney Non-CE$$p(z_{c,t}\mid z_{c,<t},u_t)=p(z_{c,t}\mid z_{c,t-1},u_t),$$4CE content CW-SSIM is $$p(z_{c,t}\mid z_{c,<t},u_t)=p(z_{c,t}\mid z_{c,t-1},u_t),$$5 and style CW-SSIM is $$p(z_{c,t}\mid z_{c,<t},u_t)=p(z_{c,t}\mid z_{c,t-1},u_t),$$6. Qualitatively, sharper anatomy and more localized enhancement are reported, while prostate enhancement is described as more diffuse and harder to generalize (Tattersall et al., 2023).

In brain tumor virtual enhancement synthesis, the dose-conditioned transformer is evaluated on BraSyn-2023 with 1,470 subjects split 1,043/219/208 for train/validation/test, using aligned T1w, T2w, FLAIR, and T1Gd plus tumor annotations. Test-set tumor-region performance for the full model reaches SSIM $$p(z_{c,t}\mid z_{c,<t},u_t)=p(z_{c,t}\mid z_{c,t-1},u_t),$$7, PSNR $$p(z_{c,t}\mid z_{c,<t},u_t)=p(z_{c,t}\mid z_{c,t-1},u_t),$$8 dB, and Dice $$p(z_{c,t}\mid z_{c,<t},u_t)=p(z_{c,t}\mid z_{c,t-1},u_t),$$9, with the best baseline MT-Net at SSIM $u_t$0, PSNR $u_t$1 dB, and Dice $u_t$2. In the tumor-region ablation, a one-step model yields SSIM $u_t$3 and PSNR $u_t$4 dB, a two-step model yields SSIM $u_t$5 and PSNR $u_t$6 dB, a model without autoregression yields SSIM $u_t$7 and PSNR $u_t$8 dB, and the full three-step model performs best (Gui et al., 2024).

In liver MRI, the multi-task T-CACE framework is evaluated on the public LLD-MMRI2023 dataset and the MG-2021 in-house dataset. On MG-2021, NCMRI$u_t$9Delay synthesis reaches MSE $u_t$0, PSNR $u_t$1 dB, SSIM $u_t$2, LPIPS $u_t$3, and FID $u_t$4, while the reported comparison to MVG is a PSNR improvement of $u_t$5 dB and an SSIM improvement of $u_t$6. The paper also reports lesion segmentation and lesion classification gains; a representative ablation row lists full T-CACE with DSC $u_t$7, IoU $u_t$8, HD95 $u_t$9, Accuracy $G$0, Sensitivity $G$1, Specificity $G$2, and F1-score $G$3. The corresponding baseline without CTE, DTAM, and time encoding is markedly lower across synthesis, segmentation, and classification metrics (Xiao et al., 13 Aug 2025).

In breast MRI, TeNCA is trained on the MAMA-MIA benchmark plus additional Duke-Breast-Cancer-MRI cases, with a test set of 300 cases, 200 patients for validation, and 1604 training cases. Images are resampled to 1 mm isotropic spacing, rescaled to $G$4, and cropped to $G$5. Test-set mean performance across all phases shows LPIPS $G$6, SSIM $G$7, MS-SSIM $G$8, PSNR $G$9 dB, FID $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$00, FRD $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$01, and parameter count $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$02. U-Net has stronger FID than the simple baseline but lower SSIM/MS-SSIM/PSNR than TeNCA, while CC-Net attains the best distribution metrics but inferior image-level metrics and is described as prone to hallucinated but pixel-wise inaccurate details (Lang et al., 23 Jun 2025).

The kinetics world model is evaluated on a private abdominal DCE-MRI dataset of 91 patients and the public Duke Breast DCE-MRI dataset of 922 exams. Using PSNR, SSIM, LPIPS, rMSE, and cSSIM, the reported abdominal results are PSNR $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$03, SSIM $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$04, LPIPS $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$05, rMSE $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$06, and cSSIM $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$07; the breast results are PSNR $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$08, SSIM $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$09, LPIPS $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$10, rMSE $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$11, and cSSIM $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$12. The combined model outperforms baseline ControlNet variants in Avg.SSIM and Avg.cSSIM, and both Latent Alignment Learning and Latent Difference Learning improve over the base model in ablation (Kong et al., 22 Feb 2026).

6. Misconceptions, limitations, and future directions

A common misconception is to equate T-CACE with a single transformer architecture. The published realizations instead include ConvLSTM-based latent recurrence, masked self-attention over decomposed tokens, phase-aware multi-task transformers, neural cellular automata, and even non-autoregressive diffusion systems that serve as precursors to explicit autoregressive extensions (Tattersall et al., 2023, Gui et al., 2024, Xiao et al., 13 Aug 2025, Lang et al., 23 Jun 2025, Kong et al., 22 Feb 2026). This suggests that the defining property of T-CACE is the temporally structured enhancement model, not a specific backbone.

A second misconception is that explicit time embeddings are always necessary. In the ConvLSTM formulation, time conditioning is implicit through sequence order and recurrent updates, and explicit embeddings were reported as unnecessary because of the strong inductive bias of convolutional recurrence (Tattersall et al., 2023). By contrast, the liver transformer uses learned phase tokens plus sinusoidal time encodings, the dose-conditioned transformer uses staircase causal masks and RoPE, the cellular automaton injects time through the update scale $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$13, and the kinetics world model uses absolute elapsed time encoded as CLIP text (Gui et al., 2024, Xiao et al., 13 Aug 2025, Lang et al., 23 Jun 2025, Kong et al., 22 Feb 2026).

A third misconception is that T-CACE is inherently physics-based. Several formulations are explicitly data-driven and do not use tracer-kinetic priors. The recurrent style-transfer model assumes the ConvLSTM can capture CE kinetics without explicit tracer-kinetic priors (Tattersall et al., 2023). TeNCA is motivated by physiological plausibility through ODE-like progression but does not hard-code Tofts-style kinetics (Lang et al., 23 Jun 2025). The kinetics world model discusses standard concentration and signal equations for context, yet its actual training objective is diffusion plus latent consistency regularization rather than explicit pharmacokinetic inversion (Kong et al., 22 Feb 2026). The liver framework injects physiological plausibility through DTAM and TCC rather than through an explicit compartment model (Xiao et al., 13 Aug 2025).

The principal limitations recur across the literature. Motion and registration remain difficult: the recurrent style-transfer work is sensitive to large inter-frame motion and does not use optical flow or registration (Tattersall et al., 2023); TeNCA notes that strong inter-phase motion or misalignment may degrade neighbor-based updates (Lang et al., 23 Jun 2025); the liver framework depends on manual multi-phase registration in its in-house dataset (Xiao et al., 13 Aug 2025); and the kinetics world model notes sensitivity to registration errors because its latent regularizers are computed from registered images (Kong et al., 22 Feb 2026). Computational cost is another concern: 3D recurrent and transformer models have substantial memory footprints, while token-based attention scales quadratically with total sequence length $$p(y_{1:T}\mid x,c)=\prod_{t=1}^{T} p(y_t\mid y_{<t},x,c),$$14 (Gui et al., 2024). Generalization under domain shift, irregular acquisition protocols, and out-of-distribution kinetics is repeatedly identified as an open issue (Xiao et al., 13 Aug 2025, Lang et al., 23 Jun 2025, Kong et al., 22 Feb 2026). Uncertainty quantification is also limited; both the liver T-CACE framework and the CAVM-derived formulation point to stochasticity or calibration as unresolved directions (Gui et al., 2024, Xiao et al., 13 Aug 2025).

Future work in this area is already sketched within the cited papers. Proposed extensions include temporal position embeddings or local causal convolutions before attention, sparse or block-diagonal attention for longer trajectories, Neural ODE-style continuous-time transitions, diffusion-augmented or hybrid decoders, evidential or Bayesian uncertainty estimation, multi-center and multi-vendor validation, explicit calibration studies, and more robust handling of misalignment and missing phases (Gui et al., 2024, Xiao et al., 13 Aug 2025, Kong et al., 22 Feb 2026). A plausible implication is that the next phase of T-CACE research will shift from demonstrating phase-consistent synthesis toward building clinically reliable generative-temporal systems that remain anatomically stable, physiologically coherent, and diagnostically calibrated under sparse and heterogeneous acquisition regimes.