Cardiac Latent Interpolation Diffusion (CaLID)

• CaLID is a generative modeling approach that utilizes latent diffusion in the VAE space to interpolate and upsample cardiac volumetric data.
• It employs geodesic latent interpolation, autoregressive bisection, and adaptive guidance to maintain anatomical fidelity and enhance reconstruction speed.
• Performance evaluations show superior metrics in PSNR, Dice, and inference time, underlining its practical utility in clinical and research cardiac imaging.

Cardiac Latent Interpolation Diffusion (CaLID) encompasses a class of generative modeling approaches that leverage latent diffusion models to solve interpolation, upsampling, and temporal synthesis tasks for cardiac volumetric data. By operating in the latent space of a variational autoencoder (VAE) or related encoding networks, these methods efficiently produce anatomically plausible intermediate volumes or slices, model spatiotemporal cardiac dynamics, and robustly reconstruct 3D or 4D (3D+t) structures from limited or sparsely sampled inputs. CaLID bridges classical geometric interpolation and registration approaches with data-driven generative modeling, enabling high-quality reconstruction, improved speed, and extensible handling of anatomic constraints in both static and dynamic cardiac imaging scenarios (Bubeck et al., 19 Aug 2025, Kim et al., 2022, Kadry et al., 2024).

1. Mathematical Foundations

CaLID methods rely fundamentally on the diffusion probabilistic model (DPM) formalism in latent space. The core mechanism views generative imaging as the denoising of a noised latent, conditioned on anatomical context.

Variational Autoencoder Backbone

For cardiac short-axis (SAX) stack interpolation (Bubeck et al., 19 Aug 2025), an encoder $E$ maps a 2D (or 3D) image $s^{(n)}$ to a latent vector $z_0$ via Gaussian parameters $(\mu_z, \Sigma_z)$, with decoder $D$ reconstructing $s^{(n)}$ from $z_0$. The VAE is trained with ELBO and adversarial LPIPS loss for anatomical detail. The latent space is typically of size $32 \times 32 \times C_{\mathrm{latent}}$ for $128 \times 128$ images.

Latent Diffusion Process

The forward latent noising process follows: $$q(z_t|z_0) = \mathcal{N}(z_t; \sqrt{\alpha_t} z_0, (1-\alpha_t)I)$$, with sequential updates $s^{(n)}$0, and the reverse process parameterized as a noise-predicting U-Net $s^{(n)}$1, leveraging context from adjacent slices or volumes. This structure follows a DDIM update scheme at inference for acceleration (Bubeck et al., 19 Aug 2025).

Interpolation and Conditioning

Missing or intermediate volumes (or slices) at position $s^{(n)}$2 are generated by sampling in the learned diffusion latent space. Contextual conditioning exploits multi-scale embeddings from neighboring image inputs, allowing spatial and spatiotemporal dependencies to be learned directly rather than imposed (Bubeck et al., 19 Aug 2025, Kim et al., 2022).

Geodesic Latent Interpolation

For continuous temporal synthesis (3D+t), latent codes are interpolated linearly (or via slerp) between endpoints, with the resulting code decoded (possibly through deformation networks) to the spatial domain, preserving the anatomical topology during motion (Kim et al., 2022, Kadry et al., 2024).

2. Architectural Components

CaLID frameworks integrate multi-component architectures tailored for cardiac data structure and the requirements of rapid, accurate reconstruction.

• VAE Encoder/Decoder: Initially trained and then frozen, with convolutional layers mapping images/volumes to compact latents.
• Diffusion Denoiser: U-Net-based, operating in low-dimensional latent space, predicting Gaussian noise at each step, with temporal or spatial context injected via embedding modules analogous to ControlNet or FiLM (Bubeck et al., 19 Aug 2025).
• Conditioning Networks: For neighboring slice context (τ₁), multi-scale feature injection (τ₂), and morpho-skeletal attributes where needed (Bubeck et al., 19 Aug 2025, Kadry et al., 2024).
• Deformation Module: In DDM (Kim et al., 2022), a VoxelMorph-style network generates a dense deformation field $s^{(n)}$3, warped through a spatial transformer to output topology-preserving deformed images.
• Guidance Mechanisms: For anatomically controlled synthesis, classifier-free, loss-based, and adaptive null guidance (ANG) strategies are applicable (Kadry et al., 2024).

Notably, all inference and interpolation is performed in the latent domain for speed and fidelity; decoders serve only to realize the final image.

3. Interpolation and Upsampling Strategies

CaLID introduces a data-driven interpolation paradigm:

• Latent Diffusion Interpolation: Instead of naive linear or spherical interpolation, CaLID leverages explicit training for interpolation in latent space using learned diffusion dynamics. Missing slices (or frames) are reconstructed by simulating the denoising process, with conditioning from adjacent acquired frames or slices (Bubeck et al., 19 Aug 2025).
• Autoregressive Bisection: Arbitrary spatial upsampling is achieved by recursively bisecting pairs of slices, generating new intermediates at $s^{(n)}$4, then further subdividing, facilitating dense 3D or temporally-resolved reconstructions (Bubeck et al., 19 Aug 2025).
• Geodesic Temporal Synthesis: For 4D (3D+t) cardiac volumes, latent trajectories interpolate motion or progression, producing intermediate states coherent with anatomic and dynamical constraints (Kim et al., 2022).
• Flexible Conditioning and Guidance: Morphological, skeletal, and other mid-level constraints can be encoded in the conditioning inputs to control the mode of interpolation or to enforce anatomical plausibility (Kadry et al., 2024).

4. Experimental Evaluation and Performance

Quantitative and qualitative experiments on CaLID and related diffusion interpolation frameworks establish state-of-the-art performance across multiple cardiac reconstruction tasks.

Metric	CaLID (DDIM 8)	DMCVR	DiffAE	DDM (Kim et al., 2022)	VoxelMorph	VM-diff
PSNR (dB)	22.75–24.16	21.02	n/a	30.73 ± 2.58	30.56	29.48
SSIM	0.755–0.778	0.631	n/a	—	—	—
Dice	0.861 (CaLID⁺)	n/a	0.823	0.802 ± 0.109	0.784	0.794
Hausdorff (mm)	2.82	n/a	n/a	—	—	—
Inference time (s)	0.125–3.0	9.0	n/a	0.456	0.219	2.90

CaLID achieves higher PSNR, SSIM, and rFID compared to DMCVR and DiffAE, with significantly reduced upsampling and generation times due to low-dimensional latent operation and accelerated DDIM sampling (8–24x speed-up) (Bubeck et al., 19 Aug 2025). In segmentation-based evaluations, CaLID's interpolated volumes improve Dice and Hausdorff Distance. In 4D modeling, DDM produces temporally and spatially coherent frames, robust topology preservation, and superior fidelity to ground truth compared to VoxelMorph and its diffeomorphic variant (Kim et al., 2022).

Annotations, segmentations, or motion fields are not required for training or inference. Robustness is observed in the context of variable slice thickness and spacing, with limitations primarily due to severe artifacts or irregular input spacings (Bubeck et al., 19 Aug 2025).

5. Extension to Spatiotemporal and Anatomic-Control Tasks

CaLID generalizes beyond 2D–3D interpolation:

• Spatiotemporal Modeling: By treating time as a third spatial coordinate and adopting 3D convolutional operations throughout encoder, decoder, and denoiser networks, CaLID reconstructs temporally coherent 2D+T (cine) or 3D+t cardiac sequences. No additional loss terms are required beyond the main diffusion objective; temporal coherence emerges from the joint spatial–temporal modeling and the conditioning architecture (Bubeck et al., 19 Aug 2025, Kim et al., 2022).
• Conditional Generation and Editing: In anatomically guided diffusion (e.g., for coronary arteries), context such as skeletons and morphometrics are injected as channels into the latent diffusion U-Net to allow morpho-skeletal manipulations, enabling the generation, interpolation, or counterfactual editing of virtual cardiac structures suitable for simulation-based device trials (Kadry et al., 2024).
• Guidance and Constraint Enforcement: Guidance strategies, particularly Adaptive Null Guidance (ANG), enable real-time enforcement of continuous anatomic constraints, facilitating controllable synthesis in clinical and simulation contexts with favorable tradeoff of fidelity, speed, and memory efficiency (Kadry et al., 2024).

6. Clinical and Procedural Implications

CaLID’s data-driven, topology-preserving strategy advances both workflow and functional assessment in cardiovascular practice and research.

• Workflow Simplification: CaLID eliminates the need for manual segmentations, explicit motion priors, or handcrafted morphometric features for whole-heart upsampling, thus streamlining standard CMR analysis pipelines (Bubeck et al., 19 Aug 2025).
• Volumetric and Functional Assessment: The improved fidelity and resolution permit more accurate calculations of ejection fraction, chamber mass, and regional wall motion from sparse CMR stacks (Bubeck et al., 19 Aug 2025).
• Virtual Interventions: The capacity for morpho-skeletal control and topology preservation enables realistic in silico experimentation, such as stent deployment simulations under varied anatomies and editing scenarios (Kadry et al., 2024).
• Extensibility: The architecture is adaptable to other modalities, including CT and ultrasound, and may generalize to broader applications in spatiotemporal medical image analysis and generative modeling (Bubeck et al., 19 Aug 2025, Kadry et al., 2024).

7. Advantages, Limitations, and Outlook

Advantages include:

• Self-supervised reconstruction (no annotation or segmentation priors) (Bubeck et al., 19 Aug 2025).
• Data-driven capture of nonlinear, anatomically faithful transitions and interpolations (Bubeck et al., 19 Aug 2025, Kim et al., 2022).
• High computational efficiency from latent space operation and few-step DDIM sampling (8–24x faster than previous SOTA) (Bubeck et al., 19 Aug 2025).
• Unified applicability to 2D→3D and 2D+T tasks with a consistent model (Bubeck et al., 19 Aug 2025).
• Robust topology preservation and spatial–temporal coherence (Kim et al., 2022).

Limitations:

• Pixel-wise fidelity may trade off against perceptual realism depending on inference step count (Bubeck et al., 19 Aug 2025).
• Sensitivity to severe input artifacts or irregular spacings (Bubeck et al., 19 Aug 2025).
• Guidance-based editing can modestly increase topology violations at high strength (Kadry et al., 2024).

A plausible implication is that further research may address enhanced handling of extreme input degradations and extend conditional guidance mechanisms to emergent anatomical attributes and modalities. The CaLID paradigm robustly advances the state of the art for cardiac volume reconstruction and generative simulation in clinical, research, and interventional contexts (Bubeck et al., 19 Aug 2025, Kim et al., 2022, Kadry et al., 2024).