Knowledge-Guided 3D CT Generation: A Conditioning-Centric Taxonomy

1 Introduction

Recent advances in deep generative modeling have enabled the synthesis of complex data across multiple domains. In medical imaging, synthetic data generation addresses limitations of real-world datasets such as high acquisition costs, annotation requirements, and privacy constraints [Koetzier et al., 2024; Lomurno and Matteucci, 2025]. Three-dimensional (3D) medical imaging modalities represent volumetric data as spatially correlated slices, offering richer anatomical information than two-dimensional imaging [Wu et al., 2025] while requiring generative models to maintain coherence across the entire volume [Friedrich et al., 2024b]. Among these, 3D computed tomography (CT) is increasingly adopted in clinical practice for diagnosing a wide range of conditions.

Early approaches to 3D CT generation focused on unconditional modeling, learning the data distribution directly from volumetric samples, but often failed to control artifacts or ensure semantic consistency [Friedrich et al., 2024b]. Supported by large-scale public CT datasets [Hamamci et al., 2024b], recent work has shifted toward knowledge-guided generative frameworks. These incorporate structured priors-semantic descriptors, anatomical constraints, or population attributes-to bias generation toward anatomically consistent, clinically meaningful volumes [Dorjsembe et al., 2024; Xu et al., 2024; Guo et al., 2025b; Yoon et al., 2025b].

This rapid growth has revealed that guidance effectiveness depends critically on knowledge type and integration mechanism [Hamamci et al., 2024a; Amirrajab et al., 2025; Molino et al., 2025]. Yet the conditioning process itself remains systematically unexamined. Existing surveys organize the field by architectural families or application domains [Zhou et al., 2025; Friedrich et al., 2024b; Chen and Ramsey, 2024; Liu et al., 2024], leaving trends, failure modes, and principled research directions obscured.

This survey addresses this gap by reframing the literature through a conditioning-centric perspective, examining how knowledge is represented, injected, and exploited throughout generation. To this end we present an innovative faceted taxonomy (Figure 1) that organizes methods along three orthogonal dimensions: (i) the type of external knowledge used for conditioning, (ii) how such knowledge is exploited within the conditioning paradigm, and (iii) the underlying generative process and model architecture. Together, these dimensions define a three-dimensional design space in which methods can be systematically positioned and compared. To summarize, the contributions of this survey are as follows:

• A conditioning-centric taxonomy for 3D CT generation, enabling principled positioning of existing methods and remaining extensible to future work.
• A unified design-space analysis grounded in the interpretable K×I×A space, revealing dominant paradigms and underexplored regions.
• Clear research directions from explicit design-space gaps, identifying actionable strategies beyond incremental refinements.

To facilitate adoption, we provide an open-source repository with reference implementations and an interactive classification tool at https://github.com/eugeniolomurno/ 3D-CT-taxonomy.

2 Preliminaries

This section formalizes knowledge-guided 3D CT generation as a constrained stochastic sampling problem over highdimensional volumetric data.

Knowledge-Guided 3D CT Generation. Let x ∈ X ⊂ RH×W×D denote a 3D CT volume represented as a voxelbased grid with inherent spatial coherence, and let k ∈ K denote external information available at generation time. Knowledge-guided CT generation is formalized as sampling from a conditional distribution pθ(x | k), with θ parameterizing the generative model and k assumed exogenous to the volumetric state. For fixed k, the conditional distribution retains stochastic support, admitting multiple plausible realizations that satisfy the imposed constraints. Knowledge-guided generation is therefore treated as constrained stochastic sampling, where modeling choices regulate how uncertainty is resolved under external constraints.

Probabilistic Conditioning. Conditioning can be instantiated through distinct probabilistic formulations. Most approaches directly parameterize pθ(x | k), treating k as a fixed input modulating the generative process during training. Others model the joint distribution pθ(x, k), where both volume and conditioning signal are co-generated, enforcing intrinsic alignment rather than treating k as an external constraint. A third paradigm modifies the sampling process at inference time: the model parameters remain fixed, but the sampling trajectory is dynamically adjusted to favor configurations consistent with k. These formulations correspond to different probabilistic objectives and encode different assumptions on how constraints interact with generative uncertainty.

Distribution Decomposition in Volumetric Generation.

The dimensionality and spatial structure of X require decomposing the target distribution into tractable components. Abstractly, volumetric synthesis proceeds by selecting a decomposition strategy for p(x | k): through global stochastic processes acting on the full volume, through explicit spatial factorizations p(x | k) = ∏n p(xn | x<n, k) as in spatialautoregressive schemes, through hierarchical multi-scale decompositions generating coarse-to-fine structure, or through deterministic invertible transformations x = T -1 (y) with y ∼ p(y | k) that redistribute spatial dependencies into structured representation spaces. In practice, these decompositions are often implemented in learned latent spaces via an encoder-decoder pair (E,D), yielding the generative objective pθ(z | k) with z = E(x) and x ≈ D(z) at reduced computational cost.

3 Taxonomy

This section introduces the proposed taxonomy for organizing 3DCT knowledge-guided approaches. In contrast to prior surveys, which predominantly organize the literature by architectural families, imaging modalities, or application domains (Table 1) [Khader et al., 2023; Friedrich et al., 2024b; Chen and Ramsey, 2024; Liu et al., 2024; Zhou et al., 2025], our formulation explicitly disentangles the conditional generative process in 3D CT synthesis. Rather than ranking methods or prescribing optimal designs, the taxonomy provides a descriptive conceptual framework for organizing and interpreting the existing literature.

3.1 Taxonomy Structure

The taxonomy is structured as a three-dimensional design space K × I × A that factorizes knowledge-guided 3D CT generation methods along three independent axes (Figure 1). Each axis captures a distinct structural aspect of model design: the type of external knowledge involved (Axis K), the mechanism by which such knowledge interacts with the generative process (Axis I), and the strategy adopted for volumetric synthesis (Axis A). Within this structure, each method is represented as a tuple (k, i, a), where k ⊆ K, i ⊆ I, and a ⊆ A. All components are non-empty, and multiple categories per axis are permitted to account for hybrid approaches. This explicit factorization enables systematic comparison across heterogeneous methods, making recurrent patterns as well as sparsely explored regions of the design space directly observable. The categories defining each axis are detailed in the following subsections, with representative instantiations summarized in Tables 2, 3, and 4.

3.2 Axis K: External Knowledge

This axis characterizes the form of external knowledge used to condition generation. Consistently with the problem formulation in Section 2, external knowledge k ∈ K is defined as information exogenous to the CT volume. Different instantiations of k vary in modality and in the degree of constraint they impose on generation. Based on these differences, we identify four broad categories, described below and illustrated with representative examples in Table 2:

K1: Textual Knowledge. K1 corresponds to unstructured semantic information encoded in linguistic form and mapped to latent representations. It provides high-level semantic conditioning without explicitly enforcing structural constraints, such that anatomical localization and organization are expected to emerge implicitly through the generative dynamics.

K2: Geometric Knowledge. K2 encodes constraints on anatomical structure and spatial organization, either through voxel-aligned representations or abstract geometric descriptors. While these priors enforce structural plausibility, they abstract away non-geometric variation, limiting the modeling of physiological patterns not directly captured by geometry.

• K3: Exemplar Knowledge. K3 comprises dense volumetric priors from reference instances. By providing voxelaligned structural and appearance information, exemplar knowledge anchors generation to specific anatomical realizations, constraining synthesis at the instance level.

K4: Attribute & Categorical Knowledge. K4 encodes non-spatial descriptive variables associated with the target volume, enabling global modulation of the data distribution. This conditioning captures population-level regularities linked to phenotypic or acquisition-related factors without prescribing local spatial structure.

Table 3: Taxonomy of Axis I (Knowledge Integration). Categories distinguish the stage and mechanism through which external knowledge influences the generative process, from pre-generative alignment to joint distribution modeling.

3.3 Axis I: Knowledge Integration

This axis characterizes how external knowledge influences the generative process, independently of its instantiation (Axis K) or the architectural backbone (Axis A). We identify four paradigms (Table 3). Throughout this section, the generative target is denoted by x, representing either volumetric data or its latent representation.

I1: Integration via Alignment. I1 paradigms enforce conditioning by establishing a correspondence between external knowledge k and x within a shared representation space prior to generation. Both modalities are mapped to compatible representations aligned under a common similarity or consistency criterion. Conditioning is thus imposed implicitly at the representation level, with knowledge influencing generation through alignment rather than direct intervention in the model's internal dynamics.

I2: Model-Based Integration. In I2 paradigm, external knowledge k is incorporated directly into the generator architecture to condition the generative dynamics, corresponding to learning a conditional predictor (e.g., ϵθ(xt, t | k)) in which knowledge modulates the model's internal computations. The integration mechanism depends on the structural properties of k: non-spatial or abstract signals induce global modulation, whereas spatially aligned knowledge enable localized conditioning that preserves correspondence. Conditioning is embedded within the learned model and consistently guides generation.

I3: Integration at Inference-Time. I3 paradigms apply conditioning exclusively at inference time, leaving the learned generative distribution unchanged. In diffusion-based models, a canonical instantiation is classifier-free guidance, which linearly combines unconditional and conditional predictions at each diffusion step t as ˜ ϵ(xt | k) = ϵθ(xt, t) + s · (˜ ϵθ(xt, t | k) - ϵθ(xt, t)), where s controls the fidelity-diversity trade-off during reverse diffusion.

I4: Integration via Joint Distribution Modeling. Unlike prior paradigms where k is fixed, I4 methods model the joint distribution pθ(x, k), generating knowledge and target variables simultaneously. This enforces intrinsic coupling between x and k, enabling jointly emerging structure and appearance. While more computationally demanding, I4 approaches remove conditioning asymmetry and capture bidirectional correlations.

3.4 Axis A: Generative Architecture

This axis characterizes the architectural strategy for volumetric synthesis, independently of knowledge representation (K) or integration (I). It describes how the high-dimensional generation problem is decomposed into tractable subproblems. We identify four strategies based on scope and representation space (Table 4).

A1: Single-Stage Latent Generation. A1 strategies model volumetric synthesis holistically by generating either the full volume x or a learned latent embedding from which x is decoded. Generation is performed in a single stage, without explicit hierarchical or spatial decomposition. This formulation favors computational efficiency and global coherence, relying on the expressive capacity of the latent representation to capture structural consistency.

A2: Multi-Stage & Cascaded Generation. A2 strategies decompose volumetric synthesis into a sequence of conditional stages, typically operating at increasing resolutions. Generation follows a coarse-to-fine hierarchy, with x(1) ∼ pθ1 (x(1) ) and x(2) ∼ pθ2 (x(2) | x(1) ), where the initial stage captures low-resolution global structure and subsequent stages progressively refine local details.

A3: Spatially-Autoregressive Generation. A3 methods factorize the generation along a spatial axis (typically the axial depth D), treating the 3D volume as a sequence of 2D slices or sub-volumes. The process follows an ordered dependency, where the generation of the n-th slice is conditioned on the previously synthesized context (x<n). This formulation models inter-slice dependencies through sequential factorization.

A4: Fixed-TransformDomain Generation. A4 strategies perform synthesis in a deterministic representation space defined by a fixed, invertible transform T . Generation is carried out over transformed coefficients y ∼ pθ(y | k), with reconstruction via x = T -1 (y). These approaches reduce dimensionality while preserving exact invertibility, at the expense of flexibility in representation learning.

4 Literature Trends

This section applies the proposed taxonomy (K×I×A) to organize and analyze the literature on knowledge-guided 3D CT generation. Table 5 provides a comprehensive overview of all reviewed methods, while Figure 2 quantifies the distribution of approaches across the three axes. The analysis is structured thematically, identifying dominant trends, recurring patterns, and underexplored regions in the proposed design space.

External Knowledge Trends. Figure 2 (top-left) shows the distribution of methods across knowledge categories. Geometric knowledge (K2) is the most prevalent, accounting for 44% of methods, instantiated primarily through organ segmentation masks. Representative approaches include MAISI, MAISI-v2, NodMAISI, LAND, and MedLoRD [Guo et al., 2025b; Zhao et al., 2025; Tushar et al., 2025; Oliveras et al., 2025; Seyfarth et al., 2025].

Exemplar conditioning (K3) accounts for 28% of existing methods and is predominantly used in cross-modal translation settings (e.g., MRI→CT). Representative approaches such as 3DLDM, 3D-WLDM, and Med-LVDM leverage source latents as structural exemplars, enabling strong alignment between input and output. Hybrid K2+K3 strategies, including DiffTumor and Lung-DDPM, demonstrate lesion-aware or layout-guided synthesis with high structural fidelity [Mahdi et al., 2025; Zheng et al., 2025; Kui et al., 2025; Chen et al., 2024; Jiang et al., 2025].

Text-based conditioning (K1) accounts for 22% of existing methods. Representative approaches include GenerateCT, Text-to-CT, Report2CT, and Text2CT, which adopt medical-specific encoders or task-adapted language models to obtain richer semantic representations [Hamamci et al., 2024a; Molino et al., 2025; Amirrajab et al., 2025; Guo et al., 2025a]. Distributional knowledge (K4) accounts for 6% of existing methods. Representative examples include Cascaded-3D and Surf2CT, as summarized in Table 5. Non-spatial descriptive variables such as demographics can be obtained without additional encoders [Yoon et al., 2025b; Yoon et al., 2025a].

Knowledge Integration Trends. The distribution across integration paradigms (Figure 2, top-center) reveals strong concentration. Model-based integration (I2) accounts for 74% of methods, incorporating external knowledge directly into the generator architecture. Representative approaches employ mechanisms such as cross-attention or feature concatenation to inject conditioning signals, enabling compatibility with all knowledge types (K1-K4) and architectural strategies. This flexibility has made I2 the most prevalent strategy for 3D CT generation [Hamamci et al., 2024a; Guo et al., 2025a; Oliveras et al., 2025; Mahdi et al., 2025; Kui et al., 2025]. Inference-time integration (I3) and joint modeling (I4) each comprise 10%. A strong coupling between K1 and I3 is observed (e.g., GenerateCT, Report2CT [Amirrajab et al., 2025], Text2CT [Guo et al., 2025a], TRACE [Shao et al., 2025]), while geometric (K2) and exemplar (K3) approaches do not adopt this paradigm. Pre-generative alignment (I1) accounts for 6% of existing methods and is currently confined to text-conditioned generation (K1), utilizing transformerbased tokenization and contrastive objectives [Hamamci et al., 2024a; Molino et al., 2025].

Architectural Trends. Figure 2 (top-center) shows that single-stage latent generation (A1) accounts for 56% of existing methods. Representative approaches such as MAISI and MAISI-v2 leverage VQ-VAE latent spaces, while LAND, MedLoRD, and DiffTumor adopt similar strategies [Guo et al., 2025b; Zhao et al., 2025; Oliveras et al., 2025; Seyfarth et al., 2025; Chen et al., 2024]. Cascaded pipelines (A2) decompose synthesis into coarse-to-fine stages and account for 15% of methods. Representative examples include GenerateCT, MedSyn, and Cascaded-3D [Hamamci et al., 2024a; Xu et al., 2024; Yoon et al., 2025b]. Autoregressive architectures (A3) represent 15% of methods and factorize generation along the axial dimension. By treating volumes as ordered slice sequences, approaches such as TRACE and CTFlow model inter-slice dependencies [Shao et al., 2025; Wang et al., 2025a]. Fixed-representation generation (A4) accounts for 12% of methods. Representative approaches include cWDM and 3D-WLDM, which operate in wavelet space [Friedrich et al., 2024a; Zheng et al., 2025].

Cross-Dimensional Patterns. The scatter plots in Figure 2 (bottom row) reveal strong dependencies between knowledge type, integration strategy, and architectural design, highlighting both consolidated practices and systematic gaps in the design space. A clear asymmetry emerges along the knowledge-integration axis (Figure 2, K-I). Textual conditioning (K1) consistently co-occurs with inference-time guidance (I3), while geometric (K2), exemplar (K3), and distributional (K4) knowledge rely almost exclusively on direct architectural integration (I2). Architectural preferences further reveal distinct patterns (Figure 2, K-A and I-A). Geometric conditioning (K2) concentrates in single-stage latent generation (A1), forming the most frequent configuration. Exemplar-based methods (K3) distribute across latent (A1) and fixed-representation architectures (A4). Textual conditioning (K1) spans latent (A1) and autoregressive designs (A3), whereas distributional knowledge (K4) appears exclusively in cascaded pipelines (A2). Model-based integration (I2) within single-stage latent models (A1) dominates the integration-architecture space. Taken together, these trends converge on the triplet (K2, I2, A1)-geometric masks with in-process modulation in single-stage latent diffusion-as the prevailing paradigm, exemplified by MAISI, MAISIv2, NodMAISI, LAND, and MedLoRD [Guo et al., 2025b; Zhao et al., 2025; Tushar et al., 2025; Oliveras et al., 2025; Seyfarth et al., 2025]. Several regions of the design space remain comparatively underrepresented.

5 Design Space Interpretation

The concentration around geometric conditioning (K2) with in-process modulation (I2) in single-stage latent diffusion (A1) reflects pragmatic convergence rather than fundamental optimality. Automated segmentation tools made voxelaligned priors widely available, while I2's architectural flexibility-enabling cross-attention, concatenation, or adaptive modulation across all knowledge types-made it universally applicable without structural constraints. Geometric masks resolve spatial uncertainty directly, enforcing anatomical correctness while constraining appearance variability. This established geometry as structural backbone, effective for targeted synthesis but limiting semantic diversity and population-aware generation. This convergence reveals three fundamental design patterns.

First, textual conditioning (K1) pairs exclusively with classifier-free guidance (I3) [Hamamci et al., 2024a; Guo et al., 2025a; Shao et al., 2025] because linguistic descriptions exhibit irreducible semantic ambiguity: for instance,

"enlarged liver" maps to diverse configurations requiring inference-time balancing between fidelity and diversity. Geometric conditioning eliminates this need by resolving spatial constraints through voxel-aligned masks, explaining why text enables semantic specification without spatial grounding while geometry ensures precision without appearance flexibility.

Second, pre-generative alignment (I1) remains confined to text-image pairs due to established vision-language frameworks (CLIP, contrastive learning), leaving geometricdemographic (K2-K4) or cross-modal exemplar (K3-K3) alignment unexplored despite potential for populationconditioned scaffolds or registration-free correspondence.

Third, architectural decompositions encode distinct tradeoffs: single-stage latent models (A1) depend critically on autoencoder fidelity, manifesting compression artifacts as slice discontinuities; cascaded pipelines (A2) propagate lowresolution errors through refinement stages; autoregressive methods (A3) accumulate sequential prediction drift; fixedtransform approaches (A4) avoid learned artifacts but cluster in MRI→CT translation where exemplar constraints (K3) stabilize wavelet-domain synthesis, is currently inapplicable to abstract conditioning (K1, K2) lacking inherent spatial structure.

Notably, some design space gaps reflect mechanistic constraints rather than oversight. Demographic attributes (K4) are structured and unambiguous (age: 65, sex: female), lacking the variability that motivates inference-time guidance (I3) for textual descriptions-direct feature modulation suffices, rendering K4-I3 combinations redundant. Similarly, fixedtransform generation (A4) with textual conditioning (K1) remains absent because linguistic semantics lack the geometric regularity required for deterministic transform domains.

6 Open Research Directions

Untapped Demographic Potential. Despite convergence on (K2, I2, A1), demographic conditioning (K4) accounts for only 6% of methods yet offers unique advantages: attributes are acquired without segmentation pipelines or learned encoders, enabling direct population-level modeling of agerelated atrophy, sex-specific morphology, or pathology prevalence at minimal annotation cost. Combining K4 with geometric scaffolds (K2) through pre-generative alignment (I1) could establish population-conditioned anatomical priors, while K4 integration via adaptive normalization (I2) could modulate global appearance without spatial constraints.

Multimodal Fusion Strategies. Multimodal integration beyond geometry remains nascent. Text-exemplar combinations (K1+K3) could ground semantic descriptions in structural priors, mitigating spatial ambiguity through voxelaligned constraints. Text-demographic pairing (K1+K4) could enable population-specific semantic generation, addressing limited spatial grounding in pure linguistic conditioning. More fundamentally, joint modeling (I4) warrants extension beyond paired image-mask generation: CT-MRI co-generation could enforce intrinsic structural alignment without explicit registration, while volume-text co-generation could prevent semantic drift inherent in conditional formulations. Combining I4 with hierarchical decompositions (A2) or fixed-transform domains (A4) could enable staged cogeneration while controlling complexity.

Architectural Diversification. Architectural exploration reveals untapped potential. Fixed-transform strategies (A4) concentrate in exemplar-driven translation but could extend to coarse geometric conditioning (K2), leveraging wavelet or spectral representations to reduce memory requirements while preserving anatomical consistency without learned compression. Conversely, identifying minimal sufficient geometric priors-treating masks as structural scaffolds while secondary sources (K1, K3, K4) modulate appearance-could substantially reduce annotation overhead. Hybrid integration strategies combining alignment (I1) with model-based conditioning (I2), or I2 with inference-time guidance (I3), remain unexplored despite complementary strengths: I1 for representation coherence, I2 for spatial modulation, I3 for generation-time controllability.

7 Limitations and Conclusions

This survey introduces the first conditioning-centric taxonomy for knowledge-guided 3D CT generation, organizing methods along orthogonal axes-knowledge type (K), integration paradigm (I), and generative architecture (A)-to enable systematic positioning within an interpretable design space. The taxonomy provides a framework for comparing conditioning strategies across heterogeneous approaches, though reliable performance assessment remains constrained by fundamental evaluation limitations.

Methods rely on non-overlapping datasets with inconsistent protocols: many are proprietary, while public data undergo arbitrary resolution adjustments, confounding conditioning contributions with dataset-specific effects. More critically, no standardized validation pipeline exists. These constraints prevent not only quantitative comparison within this survey but also reliable cross-work assessment in the literature itself-methods are rarely compared directly, obscuring which conditioning strategies, integration mechanisms, or architectural decompositions prove most effective and under what metrics. When numerical results are reported, metrics are applied non-uniformly: FID uses different feature extractors (ImageNet vs. RadImageNet), CLIP evaluations employ distinct encoders, and distributional and clinical assessments lack consensus on thresholds. Conditioning claims are often accepted without systematic ablation studies, and computational costs remain disconnected from quality improvements, preventing principled trade-off analysis.

The field's rapid evolution (2023-2025) further complicates assessment: overlapping approaches-text-based versus report-based conditioning, ControlNet versus crossattention integration-proliferate without principled comparison, obscuring incremental progress and leaving fundamental design questions unresolved. Despite these limitations, organizing literature into K×I×A design space reveals actionable opportunities and enables identification of concrete research directions, positioning future work toward cumulative refinement beyond continued proliferation of incomparable variants.

Ethical Statement

There are no ethical issues.

Acknowledgments

This paper is supported by the FAIR (Future Artificial Intelligence Research) project, funded by the NextGenerationEU program within the PNRR-PE-AI scheme (investment I.4.1) and by Fondazione Regionale per la Ricerca Biomedica (Regione Lombardia), project ID 012024R0055 PREDICT.