Paper deep dive
Knowledge-Guided 3D CT Generation: A Conditioning-Centric Taxonomy
Francesca Pia Panaccione, Eugenio Lomurno, Matteo Matteucci
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 8/13/2026, 3:47:05 AM
Summary
This paper presents a conditioning-centric taxonomy for knowledge-guided 3D Computed Tomography (CT) generation. It organizes existing methods along three orthogonal dimensions: the type of external knowledge (K), the knowledge integration paradigm (I), and the generative architecture (A). The authors analyze literature trends, identifying geometric knowledge and model-based integration as dominant, while highlighting underexplored areas to guide future research.
Entities (10)
Relation Signals (6)
Knowledge-Guided 3D CT Generation â iscategorizedby â K x I x A Taxonomy
confidence 95% ¡ In this survey, we propose a conditioning-centric taxonomy that organizes the literature along three orthogonal dimensions
I2: Model-Based Integration â isdominantin â 3D CT Generation Methods
confidence 90% ¡ Model-based integration (I2) accounts for 74% of methods
K2: Geometric Knowledge â isprevalentin â 3D CT Generation Methods
confidence 90% ¡ Geometric knowledge (K2) is the most prevalent, accounting for 44% of methods
MAISI â uses â K2: Geometric Knowledge
confidence 85% ¡ Representative approaches include MAISI... instantiated primarily through organ segmentation masks
Text-to-CT â uses â K1: Textual Knowledge
confidence 85% ¡ Representative approaches include... Text-to-CT... which adopt medical-specific encoders
DiffTumor â uses â K2: Geometric Knowledge
confidence 85% ¡ Hybrid K2+K3 strategies, including DiffTumor... demonstrate lesion-aware... synthesis
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Controllable generation guided by external knowledge is a key requirement in modern generative deep learning applications, enabling the synthesis of samples with explicit constraints on semantic content, structural properties, and variability. In 3D Computed Tomography (CT), such control is essential for clinical applications, including data augmentation, privacy-preserving data sharing, and the simulation of specific anatomical or pathological scenarios. While research on conditional 3D CT generation has expanded rapidly, the diversity of existing approaches makes systematic comparison difficult and obscures fundamental design choices. In this survey, we propose a conditioning-centric taxonomy that organizes the literature along three orthogonal dimensions: the type of external knowledge (K), the knowledge integration paradigm (I), and the generative architecture (A). This factorization defines an explicit design space (K x I x A) that provides a unified perspective on prior work. Using this framework, we systematize existing methods, identify dominant trends and recurring design patterns, and highlight underexplored regions of the design space that point toward promising directions for future research.
Tags
Links
- Source: https://arxiv.org/abs/2608.09992v1
- Canonical: https://arxiv.org/abs/2608.09992v1
Trouble viewing inline? Open PDF directly â
Full Text
47,742 characters extracted from source content.
Expand or collapse full text
Knowledge-Guided 3D CT Generation: A Conditioning-Centric Taxonomy Francesca Pia Panaccione , Eugenio Lomurno , Matteo Matteucci Department of Electronics, Information, and Bioengineering Politecnico di Milano francescapia.panaccione@polimi.it, eugenio.lomurno@polimi.it, matteo.matteucci@polimi.it Abstract Controllable generation guided by external knowl- edge is a key requirement in modern generative deep learning applications, enabling the synthesis of samples with explicit constraints on semantic content, structural properties, and variability. In 3D Computed Tomography (CT), such control is es- sential for clinical applications, including data aug- mentation, privacy-preserving data sharing, and the simulation of specific anatomical or pathological scenarios. While research on conditional 3D CT generation has expanded rapidly, the diversity of existing approaches makes systematic comparison difficult and obscures fundamental design choices. In this survey, we propose a conditioning-centric taxonomy that organizes the literature along three orthogonal dimensions: the type of external knowl- edge (K), the knowledge integration paradigm (I), and the generative architecture (A). This factoriza- tion defines an explicit design space (KĂI ĂA) that provides a unified perspective on prior work. Using this framework, we systematize existing methods, identify dominant trends and recurring design patterns, and highlight underexplored re- gions of the design space that point toward promis- ing directions for future research. 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 limi- tations 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 infor- mation 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 clini- cal practice for diagnosing a wide range of conditions. Early approaches to 3D CT generation focused on uncon- ditional modeling, learning the data distribution directly from Figure 1: Conceptual representation of the proposed taxon- omy as a three-dimensional design space. Each knowledge- guided 3D CT generation method can be positioned as a point inKĂI ĂA, defined by its choices along three orthogonal axes: external knowledge type (K), integration paradigm (I), and generative architecture (A). volumetric samples, but often failed to control artifacts or ensure semantic consistency [ Friedrich et al., 2024b ] . Sup- ported by large-scale public CT datasets [ Hamamci et al., 2024b ] , recent work has shifted toward knowledge-guided generative frameworks. These incorporate structured pri- orsâsemantic descriptors, anatomical constraints, or pop- ulation attributesâto 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 mecha- nism [ 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 arXiv:2608.09992v1 [eess.IV] 7 Aug 2026 knowledge is represented, injected, and exploited through- out generation. To this end we present an innovative faceted taxonomy (Figure 1) that organizes methods along three or- thogonal dimensions: (i) the type of external knowledge used for conditioning, (i) how such knowledge is exploited within the conditioning paradigm, and (i) 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 summa- rize, the contributions of this survey are as follows: ⢠A conditioning-centric taxonomy for 3D CT genera- tion, enabling principled positioning of existing methods and remaining extensible to future work. ⢠A unified design-space analysis grounded in the inter- pretableKĂIĂA space, revealing dominant paradigms and underexplored regions. ⢠Clear research directions from explicit design-space gaps, identifying actionable strategies beyond incremen- tal refinements. To facilitate adoption, we provide an open-source repos- itory with reference implementations and an interac- tive classification tool at https://github.com/eugeniolomurno/ 3D-CT-taxonomy. 2 Preliminaries This section formalizes knowledge-guided 3D CT genera- tion as a constrained stochastic sampling problem over high- dimensional volumetric data. Knowledge-Guided 3D CT Generation. Let x â X â R HĂWĂD denote a 3D CT volume represented as a voxel- based grid with inherent spatial coherence, and let k â K denote external information available at generation time. Knowledge-guided CT generation is formalized as sampling from a conditional distribution p θ (x | k), with θ parameter- izing the generative model and k assumed exogenous to the volumetric state. For fixed k, the conditional distribution re- tains stochastic support, admitting multiple plausible realiza- tions that satisfy the imposed constraints. Knowledge-guided generation is therefore treated as constrained stochastic sam- pling, where modeling choices regulate how uncertainty is resolved under external constraints. Probabilistic Conditioning. Conditioning can be instanti- ated through distinct probabilistic formulations. Most ap- proaches directly parameterize p θ (x | k), treating k as a fixed input modulating the generative process during train- ing. Others model the joint distribution p θ (x,k), where both volume and conditioning signal are co-generated, enforcing intrinsic alignment rather than treating k as an external con- straint. A third paradigm modifies the sampling process at inference time: the model parameters remain fixed, but the sampling trajectory is dynamically adjusted to favor config- urations consistent with k. These formulations correspond to different probabilistic objectives and encode different as- sumptions on how constraints interact with generative uncer- tainty. SurveyOrganization BreadthQnt. Analysis Design Space [ Khader et al., 2023 ] ArchitectureMed-3DPerformanceâ [ Friedrich et al., 2024b ] ModalityMed-3Dâ [ Liu et al., 2024 ] ApplicationBrain/Heart Performanceâ [ Chen and Ramsey, 2024 ] ArchitectureGen-3Dâ [ Zhou et al., 2025 ] ApplicationMed-AllPerformanceâ OursConditioning3D CTDistributionalâ Table 1: Comparison with related surveys. This work in- troduces the first conditioning-centric taxonomy with explicit method positioning a structured design space (KĂIĂA). Distribution Decomposition in Volumetric Generation. The dimensionality and spatial structure of X require de- composing the target distribution into tractable components. Abstractly, volumetric synthesis proceeds by selecting a de- composition strategy for p(x | k): through global stochastic processes acting on the full volume, through explicit spatial factorizations p(x | k) = Q n p(x n | x <n ,k) as in spatial- autoregressive schemes, through hierarchical multi-scale de- compositions generating coarse-to-fine structure, or through deterministic invertible transformations x = T â1 (y) with y âź p(y | k) that redistribute spatial dependencies into struc- tured representation spaces. In practice, these decomposi- tions are often implemented in learned latent spaces via an encoderâdecoder pair (E,D), yielding the generative objec- tive p θ (z | k) with z = E(x) and x â D(z) at reduced computational cost. 3 Taxonomy This section introduces the proposed taxonomy for organiz- ing 3D CT knowledge-guided approaches. In contrast to prior surveys, which predominantly organize the literature by ar- chitectural families, imaging modalities, or application do- mains (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 gener- ative process in 3D CT synthesis. Rather than ranking meth- ods or prescribing optimal designs, the taxonomy provides a descriptive conceptual framework for organizing and inter- preting the existing literature. 3.1 Taxonomy Structure The taxonomy is structured as a three-dimensional design space KĂI ĂA that factorizes knowledge-guided 3D CT generation methods along three independent axes (Figure 1). Each axis captures a distinct structural aspect of model de- sign: the type of external knowledge involved (Axis K), the mechanism by which such knowledge interacts with the gen- erative process (Axis I), and the strategy adopted for volu- metric synthesis (AxisA). Within this structure, each method is represented as a tuple (k,i,a), where k â K, i â I, and a â A. All components are non-empty, and multiple categories per axis are permitted to account for hybrid ap- proaches. This explicit factorization enables systematic com- parison across heterogeneous methods, making recurrent pat- terns as well as sparsely explored regions of the design space directly observable. The categories defining each axis are de- Axis K: External Knowledge CategoryRepresentative Instantiations K1 - Textual Knowledge Radiology reports, free-text clinical de- scriptions, open-vocabulary prompts, un- structured findings. K2 - Geometric Knowledge Organ segmentation masks, anatomical layouts, bounding boxes, landmark coordi- nates, sparse structural maps. K3 - Exemplar Knowledge Reference volumes from complementary modalities (e.g., MRI, PET, CBCT), patient-specific priors, atlas templates. K4 - Attribute & Categorical Knowledge Patient demographics (e.g., age, sex, BMI), diagnostic class labels, acquisition metadata, structured clinical attributes. Table 2: Taxonomy of Axis K (External Knowledge). Cat- egories represent different types of information used to con- dition 3D CT generation, ranging from unstructured text to dense volumetric priors. tailed in the following subsections, with representative instan- tiations summarized in Tables 2, 3, and 4. 3.2 AxisK: External Knowledge This axis characterizes the form of external knowledge used to condition generation. Consistently with the problem for- mulation in Section 2, external knowledge k â K is defined as information exogenous to the CT volume. Different instan- tiations 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 con- ditioning without explicitly enforcing structural constraints, such that anatomical localization and organization are ex- pected 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 descrip- tors. 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 volumet- ric priors from reference instances. By providing voxel- aligned structural and appearance information, exemplar knowledge anchors generation to specific anatomical realiza- tions, 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 distribu- tion. This conditioning captures population-level regulari- ties linked to phenotypic or acquisition-related factors with- out prescribing local spatial structure. Axis I: Knowledge Integration CategoryRepresentative Instantiations I1 - Integration via Alignment Pre-generative embedding alignment (e.g., CLIP, BioViL), dual-encoder architec- tures, contrastive representation learning. I2 - Model-Based Integration Cross-attentionlayers,channel-wise concatenation, feature-wise modulation (FiLM, AdaGN), ControlNet adapters. I3 - Integration at Inference-Time Classifier-free guidance,energy-based Guidance, trajectory smoothing, composi- tional guidance. I4 - Integration via Joint distribution modeling joint probability factorization p(x, k), uni- fied diffusion over concatenated modali- ties (e.g., image + mask). Table 3: Taxonomy of AxisI (Knowledge Integration). Cat- egories distinguish the stage and mechanism through which external knowledge influences the generative process, from pre-generative alignment to joint distribution modeling. 3.3 AxisI: Knowledge Integration This axis characterizes how external knowledge influences the generative process, independently of its instantiation (AxisK) or the architectural backbone (AxisA). We identify four paradigms (Table 3). Throughout this section, the gen- erative target is denoted by x, representing either volumetric data or its latent representation. I1: Integration via Alignment. I1 paradigms enforce con- ditioning 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 con- sistency criterion. Conditioning is thus imposed implicitly at the representation level, with knowledge influencing genera- tion through alignment rather than direct intervention in the modelâs internal dynamics. I2: Model-Based Integration. In I2 paradigm, external knowledge k is incorporated directly into the generator ar- chitecture to condition the generative dynamics, correspond- ing to learning a conditional predictor (e.g., Îľ θ (x t ,t | k)) in which knowledge modulates the modelâs internal computa- tions. The integration mechanism depends on the structural properties of k: non-spatial or abstract signals induce global modulation, whereas spatially aligned knowledge enable lo- calized conditioning that preserves correspondence. Condi- tioning is embedded within the learned model and consis- tently guides generation. I3:Integration at Inference-Time. I3 paradigms ap- ply 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 pre- dictions at each diffusion step t as ĚÎľ(x t | k) = Îľ θ (x t ,t) + s ¡ ( Ěξθ(x t ,t | k) â ξθ(x t ,t)), where s controls the fi- delityâdiversity trade-off during reverse diffusion. Axis A: Generative Architecture CategoryRepresentative Instantiations A1 - Single-Stage Latent Generation Holistic generation in learned latent spaces, 3D-VQGAN backbones, one-shot synthesis. A2 - Multi-Stage & Cascaded Generation Coarse-to-fine synthesis, super-resolution cascades, hierarchical residual modeling. A3 - Spatially- Autoregressive Generation Slice-wise autoregression, video-like gen- eration (z-axis evolution), temporal trans- formers, spatial decomposition. A4 - Fixed-Transform Domain Wavelet-domain diffusion, spectral diffu- sion, generation in deterministic invertible spaces. Table 4: Taxonomy of Axis A (Generative Architecture). Categories characterize the structural strategy for volumetric synthesis, distinguishing methods by their procedural decom- position and representation space. I4: Integration via Joint Distribution Modeling. Unlike prior paradigms where k is fixed, I4 methods model the joint distribution p θ (x,k), generating knowledge and target vari- ables simultaneously. This enforces intrinsic coupling be- tween x and k, enabling jointly emerging structure and ap- pearance. While more computationally demanding, I4 ap- proaches remove conditioning asymmetry and capture bidi- rectional correlations. 3.4 AxisA: Generative Architecture This axis characterizes the architectural strategy for volumet- ric synthesis, independently of knowledge representation (K) or integration (I). It describes how the high-dimensional gen- eration 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 de- coded. Generation is performed in a single stage, without explicit hierarchical or spatial decomposition. This formu- lation 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 condi- tional stages, typically operating at increasing resolutions. Generation follows a coarse-to-fine hierarchy, with x (1) âź p θ 1 (x (1) ) and x (2) âź p θ 2 (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 ax- ial depth D), treating the 3D volume as a sequence of 2D slices or sub-volumes. The process follows an ordered depen- dency, where the generation of then-th slice is conditioned on the previously synthesized context (x <n ). This formulation models inter-slice dependencies through sequential factoriza- tion. A4: Fixed-Transform Domain Generation. A4 strategies perform synthesis in a deterministic representation space de- fined by a fixed, invertible transform T . Generation is carried out over transformed coefficients y âź p θ (y | k), with recon- struction via x = T â1 (y). These approaches reduce dimen- sionality while preserving exact invertibility, at the expense of flexibility in representation learning. 4 Literature Trends This section applies the proposed taxonomy (KĂIĂA) to or- ganize 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. Geo- metric knowledge (K2) is the most prevalent, accounting for 44% of methods, instantiated primarily through organ seg- mentation 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., MRIâCT). Representative approaches such as 3DLDM, 3D-WLDM, and Med-LVDM leverage source la- tents as structural exemplars, enabling strong alignment be- tween 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 ex- isting methods. Representative approaches include Gener- ateCT, Text-to-CT, Report2CT, and Text2CT, which adopt medical-specific encoders or task-adapted language mod- els 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 in- clude 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 con- catenation to inject conditioning signals, enabling compat- ibility with all knowledge types (K1âK4) and architectural strategies. This flexibility has made I2 the most prevalent MethodKIAConditioning Source Conditioning MechanismGeneration StrategyDataset (Resolution)Code MedGen3D [ Han et al., 2023 ] K2I4A3Voxel-aligned Semantic Maps Joint maskâimage diffusion (MC-DPM) Autoregressive slice-wise SegTHOR (96Ă320 2 )â GenerateCT [ Hamamci et al., 2024a ] K1I1, I2A2Free-form Text Prompts Cross-attention, CFGCascaded low-resâsup-res CT-RATE (512 2 Ă201)§ MedSyn [ Xu et al., 2024 ] K1, K2I4A2Textual Description + Anatomical Semantic Layouts Joint DiffusionMulti-Staged low-resâsup-res Private Lung Dataset (256 3 )§ GEM-3D [ Zhu et al., 2024 ] K2, K3I2A3Anatomical Masks + Reference slice Latent ConcatenationSequential Window-based AbdomenCT-1K (512 2 ĂZ, Z variable) § CM3dLDM [ Tapp et al., 2024 ] K3I2A1Cross-modal Volume (MRI) Frozen Encoder InjectionSingle-stage Patch-based Diffusion Private MR-CT Head Dataset; SynthRad (224 3 ) § cWDM [ Friedrich et al., 2024a ] K3I2A4Cross-modal Volume (MRI) Wavelet-domain Feature Modulation Fixed-transform Wavelet Diffusion BraTS 2024 (128 3 )§ DiffTumor [ Chen et al., 2024 ] K2, K3I2A1Masks + Healthy CT Volume Latent Patch ConcatenationSingle-stage Latent Diffusion (Inpainting) LiTS; MSD; KiTS; AbdAtlas-8K; Hopkins (96 3 ) § MC-IDDDPM [ Pan et al., 2024 ] K3I2A4Cross-modal Volume (MRI) Feature ConcatenationSingle-stage DiffusionPrivate Brain and Prostate Datasets MRI-CT (192 2 Ă96) â LN-DDPM [ Yu et al., 2024 ] K2I2A1Lymph Node + Organ Masks Spatial ConcatenationSingle-stage patch-based diffusion Private Colorectal Dataset; ABD-LN; (128 3 ) â Med-LVDM [ Kui et al., 2025 ] K3I2A1Cross-modal Volume (MRI) Latent concatenationSingle-stage Latent Diffusion Pelvis MRâCT (256 2 ĂZ)§ Cascaded-3D [ Yoon et al., 2025b ] K4I2A2DemographicsAdaGNCascaded coarseâ sup-res AutoPET (224Ă224Ă384)â Surf2CT [ Yoon et al., 2025a ] K2,K4I2A2Skin Surface + Demographics Conditional flow matchingCascaded coarseâsuper-res Private Torso CT Dataset; AutoPET (224Ă224Ă352) â CTFlow [ Wang et al., 2025a ] K1I2A3Clinical ReportsCross-attentionSlice-as-video Flow Matching CT-RATE (256 2 ĂZ, Z variable)â Report2CT [ Amirrajab et al., 2025 ] K1I2, I3A1Clinical ReportsMulti-encoder Cross-Attention, CFG Single-stage Latent Diffusion CT-RATE (480 2 Ă256)§ Text-to-CT [ Molino et al., 2025 ] K1I1, I2A1Radiology text3D CLIP pretraining, Cross-Attention Single-stage Latent Diffusion CT-RATE (512 2 Ă128)§ Text2CT [ Guo et al., 2025a ] K1I2, I3A1Free-text descriptions Cross-attention, CFGSingle-stage Latent Diffusion CT-RATE; RadChestCT (512 2 Ă192) â LAND [ Oliveras et al., 2025 ] K2I2A1Lung + Nodule Segmentation Masks Latent Concatenation, Cross Attention Single-stage Latent Diffusion LIDC-IDRI; NLST (256 3 )â MedLoRD [ Seyfarth et al., 2025 ] K2I2A1Anatomical Segmentation Masks ControlNetSingle-stage Latent Diffusion Private Coronary Dataset; LUNA16 (512 2 Ă256) § MAISI [ Guo et al., 2025b ] K2I2A1Multi-organ Segmentation Maps ControlNetSingle-stage Latent Diffusion Curated multi-source, multi-organ CT dataset (512 2 Ă768) § MAISI-v2 [ Zhao et al., 2025 ] K2I2A1Multi-organ Segmentation Maps ControlNet, RCLSingle-stage latentCurated multi-source multi-organ CT dataset (512 2 Ă768) § NodMAISI [ Tushar et al., 2025 ] K2I2A1Nodule-specific Semantic Masks ControlNetSingle-stage Latent Diffusion Curated multi-source public lung CT datasets (512 2 Ă768) § TRACE [ Shao et al., 2025 ] K1, K2I2, I3A3Text + Multi-modal Anatomical Masks Multi-modal Mask Concatenation Autoregressive Slice-pair (Video-like) CT-RATE (256 2 ĂZ, Z variable)§ Lung-DDPM [ Jiang et al., 2025 ] K2, K3I2A1Semantic Layouts + Reference CT Layout concatenation + AASSingle-stage Layout-guided Private Lung Dataset; LIDC-IDRI (128 3 ) § 3DLDM [ Mahdi et al., 2025 ] K3I2A1Cross-modal Volume (MRI) Latent Feature ConcatenationSingle-stage Latent Diffusion SynthRAD2023 (96 2 Ă256)â 3D-WLDM [ Zheng et al., 2025 ] K3I2A4Cross-modal Volume (MRI) Latent Wavelet ConcatenationFixed-transform Latent Diffusion Private PET/MRâPET/CT Datasets (128 3 ) â LabelG [ Wang et al., 2025b ] K2I4A1Semantic Segmentation Masks Joint Latent-space ModelingSingle-stage Latent Diffusion Private Abdominal CT dataset, AbdomenCT; SegTHOR; MSD10-Colon; (256Ă256Ă128) â Table 5: Comprehensive overview of knowledge-guided 3D CT generation methods organized under the proposed taxonomy. For each method, we report the classification along axes K, I, and A, the conditioning source, the conditioning mechanism, the generation strategy, the dataset used with volumetric resolution (as reported by authors when available), and code availability. Methods using multiple categories within a single axis are listed with all applicable labels. Note: CFG = Classifier-Free Guidance; AdaGN = Adaptive Group Normalization; RCL = Region-specific Contrastive Loss; AAS = Anatomically-Aware Sampling; MC-DPM = Multi-Condition Denoising Probabilistic Model.§ indicates publicly available code repository. 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) ac- counts for 6% of existing methods and is currently confined to text-conditioned generation (K1), utilizing transformer- based 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 ex- Figure 2: Statistical distribution of methods across the taxonomy. Top row: frequency of each category in axesK, I, andA, plus single versus multiple category usage per axis. Bottom row: pairwise interactions between axes, where node size indicates method count and pie slices show the third axisâs distribution (methods using multiple categories appear in multiple counts). isting 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) de- compose synthesis into coarse-to-fine stages and account for 15% of methods. Representative examples include Genera- teCT, MedSyn, and Cascaded-3D [ Hamamci et al., 2024a; Xu et al., 2024; Yoon et al., 2025b ] . Autoregressive archi- tectures (A3) represent 15% of methods and factorize gen- eration 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 knowl- edge type, integration strategy, and architectural design, high- lighting both consolidated practices and systematic gaps in the design space. A clear asymmetry emerges along the knowledgeâintegration axis (Figure 2, KâI). Textual condi- tioning (K1) consistently co-occurs with inference-time guid- ance (I3), while geometric (K2), exemplar (K3), and distri- butional (K4) knowledge rely almost exclusively on direct architectural integration (I2). Architectural preferences fur- ther reveal distinct patterns (Figure 2, KâA and IâA). Geo- metric 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 condi- tioning (K1) spans latent (A1) and autoregressive designs (A3), whereas distributional knowledge (K4) appears exclu- sively in cascaded pipelines (A2). Model-based integration (I2) within single-stage latent models (A1) dominates the integrationâarchitecture space. Taken together, these trends converge on the triplet (K2, I2, A1)âgeometric masks with in-process modulation in single-stage latent diffusionâas the prevailing paradigm, exemplified by 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 ] . 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 fundamen- tal optimality. Automated segmentation tools made voxel- aligned priors widely available, while I2âs architectural flex- ibilityâenabling cross-attention, concatenation, or adaptive modulation across all knowledge typesâmade it univer- sally applicable without structural constraints. Geometric masks resolve spatial uncertainty directly, enforcing anatom- ical 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 descrip- tions exhibit irreducible semantic ambiguity: for instance, "enlarged liver" maps to diverse configurations requiring inference-time balancing between fidelity and diversity. Ge- ometric conditioning eliminates this need by resolving spa- tial constraints through voxel-aligned masks, explaining why text enables semantic specification without spatial grounding while geometry ensures precision without appearance flexi- bility. Second, pre-generative alignment (I1) remains confined to text-image pairs due to established vision-language frameworks (CLIP, contrastive learning), leaving geometric- demographic (K2âK4) or cross-modal exemplar (K3âK3) alignment unexplored despite potential for population- conditioned scaffolds or registration-free correspondence. Third, architectural decompositions encode distinct trade- offs: single-stage latent models (A1) depend critically on autoencoder fidelity, manifesting compression artifacts as slice discontinuities; cascaded pipelines (A2) propagate low- resolution errors through refinement stages; autoregressive methods (A3) accumulate sequential prediction drift; fixed- transform approaches (A4) avoid learned artifacts but cluster in MRIâCT translation where exemplar constraints (K3) sta- bilize wavelet-domain synthesis, is currently inapplicable to abstract conditioning (K1, K2) lacking inherent spatial struc- ture. Notably, some design space gaps reflect mechanistic con- straints rather than oversight. Demographic attributes (K4) are structured and unambiguous (age: 65, sex: female), lack- ing the variability that motivates inference-time guidance (I3) for textual descriptionsâdirect feature modulation suffices, rendering K4-I3 combinations redundant. Similarly, fixed- transform generation (A4) with textual conditioning (K1) re- mains 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 en- coders, enabling direct population-level modeling of age- related atrophy, sex-specific morphology, or pathology preva- lence at minimal annotation cost.Combining K4 with geometric scaffolds (K2) through pre-generative alignment (I1) could establish population-conditioned anatomical pri- ors, 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 com- binations (K1+K3) could ground semantic descriptions in structural priors, mitigating spatial ambiguity through voxel- aligned constraints.Text-demographic pairing (K1+K4) could enable population-specific semantic generation, ad- dressing limited spatial grounding in pure linguistic condi- tioning. 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 formula- tions. Combining I4 with hierarchical decompositions (A2) or fixed-transform domains (A4) could enable staged co- generation while controlling complexity. Architectural Diversification. Architectural exploration reveals untapped potential. Fixed-transform strategies (A4) concentrate in exemplar-driven translation but could ex- tend to coarse geometric conditioning (K2), leveraging wavelet or spectral representations to reduce memory re- quirements while preserving anatomical consistency without learned compression. Conversely, identifying minimal suf- ficient geometric priorsâtreating masks as structural scaf- folds while secondary sources (K1, K3, K4) modulate ap- pearanceâcould 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 mod- ulation, I3 for generation-time controllability. 7 Limitations and Conclusions This survey introduces the first conditioning-centric taxon- omy for knowledge-guided 3D CT generation, organizing methods along orthogonal axesâknowledge type (K), inte- gration paradigm (I), and generative architecture (A)âto en- able systematic positioning within an interpretable design space. The taxonomy provides a framework for compar- ing conditioning strategies across heterogeneous approaches, though reliable performance assessment remains constrained by fundamental evaluation limitations. Methods rely on non-overlapping datasets with inconsis- tent protocols: many are proprietary, while public data un- dergo arbitrary resolution adjustments, confounding condi- tioning contributions with dataset-specific effects. More crit- ically, no standardized validation pipeline exists. These con- straints prevent not only quantitative comparison within this survey but also reliable cross-work assessment in the liter- ature itselfâmethods are rarely compared directly, obscur- ing 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 extrac- tors (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 computa- tional costs remain disconnected from quality improvements, preventing principled trade-off analysis. The fieldâs rapid evolution (2023â2025) further compli- cates assessment: overlapping approachesâtext-based ver- sus report-based conditioning, ControlNet versus cross- attention integrationâproliferate without principled compar- ison, obscuring incremental progress and leaving fundamen- tal design questions unresolved. Despite these limitations, organizing literature into KĂIĂA design space reveals ac- tionable opportunities and enables identification of concrete research directions, positioning future work toward cumula- tive refinement beyond continued proliferation of incompara- ble variants. Ethical Statement There are no ethical issues. Acknowledgments This paper is supported by the FAIR (Future Artificial Intel- ligence 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 (Re- gione Lombardia), project ID 012024R0055 PREDICT. References [ Amirrajab et al., 2025 ] Sina Amirrajab, Zohaib Salahuddin, Sheng Kuang, Henry C Woodruff, and Philippe Lam- bin. Radiology report conditional 3d ct generation with multi encoder latent diffusion model.arXiv preprint arXiv:2509.14780, 2025. [ Chen and Ramsey, 2024 ] Kaiqi Chen and Libby Ramsey. Deep generative models for 3d content creation: A com- prehensive survey of architectures, challenges, and emerg- ing trends. 2024. [ Chen et al., 2024 ] Qi Chen, Xiaoxi Chen, Haorui Song, Zhiwei Xiong, Alan Yuille, Chen Wei, and Zongwei Zhou. Towards generalizable tumor synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11147â11158, 2024. [ Dorjsembe et al., 2024 ] Zolnamar Dorjsembe, Hsing-Kuo Pao, Sodtavilan Odonchimed, and Furen Xiao. Condi- tional diffusion models for semantic 3d brain mri synthe- sis. IEEE Journal of Biomedical and Health Informatics, 28(7):4084â4093, 2024. [ Friedrich et al., 2024a ] Paul Friedrich, Alicia Durrer, Ju- lia Wolleb, and Philippe C Cattin. cwdm: Conditional wavelet diffusion models for cross-modality 3d medical image synthesis. arXiv preprint arXiv:2411.17203, 2024. [ Friedrich et al., 2024b ] Paul Friedrich, Yannik Frisch, and Philippe C Cattin. Deep generative models for 3d medical image synthesis. In Generative Machine Learning Models in Medical Image Computing, pages 255â278. Springer, 2024. [ Guo et al., 2025a ] Pengfei Guo, Can Zhao, Dong Yang, Yu- fan He, Vishwesh Nath, Ziyue Xu, Pedro RAS Bassi, Zongwei Zhou, Benjamin D Simon, Stephanie Anne Har- mon, et al. Text2ct: Towards 3d ct volume generation from free-text descriptions using diffusion model. arXiv preprint arXiv:2505.04522, 2025. [ Guo et al., 2025b ] Pengfei Guo, Can Zhao, Dong Yang, Ziyue Xu, Vishwesh Nath, Yucheng Tang, Benjamin Si- mon, Mason Belue, Stephanie Harmon, Baris Turkbey, et al. Maisi: Medical ai for synthetic imaging. In 2025 IEEE/CVF Winter Conference on Applications of Com- puter Vision (WACV), pages 4430â4441. IEEE, 2025. [ Hamamci et al., 2024a ] Ibrahim Ethem Hamamci, Sezgin Er, Anjany Sekuboyina, Enis Simsar, Alperen Tezcan, Ayse Gulnihan Simsek, Sevval Nil Esirgun, Furkan Almas, Irem Do Ě gan, Muhammed Furkan Dasdelen, et al. Gener- atect: Text-conditional generation of 3d chest ct volumes. In European Conference on Computer Vision, pages 126â 143. Springer, 2024. [ Hamamci et al., 2024b ] Ibrahim Ethem Hamamci, Sezgin Er, Chenyu Wang, Furkan Almas, Ayse Gulnihan Simsek, Sevval Nil Esirgun, Irem Dogan, Omer Faruk Durugol, Benjamin Hou, Suprosanna Shit, et al. Developing gener- alist foundation models from a multimodal dataset for 3d computed tomography. arXiv preprint arXiv:2403.17834, 2024. [ Han et al., 2023 ] Kun Han, Yifeng Xiong, Chenyu You, Pooya Khosravi, Shanlin Sun, Xiangyi Yan, James S Duncan, and Xiaohui Xie. Medgen3d: A deep genera- tive framework for paired 3d image and mask generation. In International Conference on Medical Image Comput- ing and Computer-Assisted Intervention, pages 759â769. Springer, 2023. [ Jiang et al., 2025 ] Yifan Jiang, Yannick LemarĂŠchal, JosĂŠe Bafaro, Jessica Abi-Rjeile, Philippe Joubert, Philippe De- sprĂŠs, and Venkata Manem. Lung-ddpm: Semantic layout- guided diffusion models for thoracic ct image synthesis. arXiv preprint arXiv:2502.15204, 2025. [ Khader et al., 2023 ] Firas Khader, Gustav MĂźller-Franzes, Soroosh Tayebi Arasteh, Tianyu Han, Christoph Haar- burger, Maximilian Schulze-Hagen, Philipp Schad, Sandy Engelhardt, Bettina BaeĂler, Sebastian Foersch, et al. De- noising diffusion probabilistic models for 3d medical im- age generation. Scientific Reports, 13(1):7303, 2023. [ Koetzier et al., 2024 ] LennartRKoetzier,JieWu, Domenico Mastrodicasa, Aline Lutz, Matthew Chung, W Adam Koszek, Jayanth Pratap, Akshay S Chaudhari, Pranav Rajpurkar, Matthew P Lungren, et al.Gener- ating synthetic data for medical imaging.Radiology, 312(3):e232471, 2024. [ Kui et al., 2025 ] Xiaoyan Kui, Bo Liu, Zanbo Sun, Qinsong Li, Min Zhang, Wei Liang, and Beiji Zou. Med-lvdm: Medical latent variational diffusion model for medical im- age translation. Biomedical Signal Processing and Con- trol, 106:107735, 2025. [ Liu et al., 2024 ] Yanbin Liu, Girish Dwivedi, Farid Bous- said, and Mohammed Bennamoun. 3d brain and heart vol- ume generative models: a survey. ACM Computing Sur- veys, 56(6):1â37, 2024. [ Lomurno and Matteucci, 2025 ] Eugenio Lomurno and Mat- teo Matteucci. Federated knowledge recycling: Privacy- preserving synthetic data sharing. Pattern Recognition Letters, 191:124â130, 2025. [ Mahdi et al., 2025 ] Mohammed A Mahdi, Mohammed Al- Shalabi, Ehab T Alnfrawy, Reda Elbarougy, Muham- mad Usman Hadi, and Rao Faizan Ali. 3d latent diffusion model for mr-only radiotherapy: Accurate and consistent synthetic ct generation. Diagnostics, 15(23):3010, 2025. [ Molino et al., 2025 ] DanieleMolino,CamilloMaria Caruso, Filippo Ruffini, Paolo Soda, and Valerio Guarrasi. Text-to-ct generation via 3d latent diffusion model with contrastive vision-language pretraining.arXiv preprint arXiv:2506.00633, 2025. [ Oliveras et al., 2025 ] Anna Oliveras, Roger MarĂ, Rafael Redondo, Oriol GuardiĂ , Ana Tost, Bhalaji Nagarajan, Carolina Migliorelli, Vicent Ribas, and Petia Radeva. Land:Lung and nodule diffusion for 3d chest ct synthesis with anatomical guidance.arXiv preprint arXiv:2510.18446, 2025. [ Pan et al., 2024 ] Shaoyan Pan,Elham Abouei,Jacob Wynne, Chih-Wei Chang, Tonghe Wang, Richard LJ Qiu, Yuheng Li, Junbo Peng, Justin Roper, Pretesh Pa- tel, et al.Synthetic ct generation from mri using 3d transformer-based denoising diffusion model.Medical Physics, 51(4):2538â2548, 2024. [ Seyfarth et al., 2025 ] Marvin Seyfarth, Salman Ul Hassan Dar, Isabelle Ayx, Matthias Alexander Fink, Stefan O Schoenberg, Hans-Ulrich Kauczor, and Sandy Engelhardt. Medlord: A medical low-resource diffusion model for high-resolution 3d ct image synthesis. In International Workshop on Simulation and Synthesis in Medical Imag- ing, pages 1â12. Springer, 2025. [ Shao et al., 2025 ] Minye Shao, Xingyu Miao, Haoran Duan, Zeyu Wang, Jingkun Chen, Yawen Huang, Xian Wu, Jingjing Deng, Yang Long, and Yefeng Zheng. Trace: Temporally reliable anatomically-conditioned 3d ct gen- eration with enhanced efficiency. In International Con- ference on Medical Image Computing and Computer- Assisted Intervention, pages 627â637. Springer, 2025. [ Tapp et al., 2024 ] Austin Tapp, Abhijeet Parida, Can Zhao, Van Lam, Natasha Lepore, Syed Muhammad Anwar, and Marius George Linguraru. Mr to ct synthesis using 3d la- tent diffusion. In 2024 IEEE International Symposium on Biomedical Imaging (ISBI), pages 1â5. IEEE, 2024. [ Tushar et al., 2025 ] Fakrul Islam Tushar, Ehsan Samei, Cynthia Rudin, and Joseph Y Lo. Nodmaisi: Nodule- oriented medical ai for synthetic imaging. arXiv preprint arXiv:2512.18038, 2025. [ Wang et al., 2025a ] Jiayi Wang, Hadrien Reynaud, Fran- ciskus Xaverius Erick, and Bernhard Kainz.Ctflow: Video-inspired latent flow matching for 3d ct synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6750â6758, 2025. [ Wang et al., 2025b ] Lu-Yan Wang, Tzung-Dau Wang, and Shang-Hong Lai. Labelg : Consistent pairwise 3d CT im- age and segmentation mask generation via medical foun- dation model. In Submitted to Medical Imaging with Deep Learning, 2025. under review. [ Wu et al., 2025 ] Jing Wu, Yuli Wang, Zhusi Zhong, Weihua Liao, Natalia Trayanova, Zhicheng Jiao, and Harrison X Bai. Vision-language foundation model for 3d medical imaging. npj Artificial Intelligence, 1(1):17, 2025. [ Xu et al., 2024 ] Yanwu Xu, Li Sun, Wei Peng, Shuyue Jia, Katelyn Morrison, Adam Perer, Afrooz Zandifar, Shyam Visweswaran, Motahhare Eslami, and Kayhan Bat- manghelich. Medsyn: text-guided anatomy-aware synthe- sis of high-fidelity 3-d ct images. IEEE Transactions on Medical Imaging, 43(10):3648â3660, 2024. [ Yoon et al., 2025a ] Siyeop Yoon, Yujin Oh, Pengfei Jin, Sifan Song, Matthew Tivnan, Dufan Wu, Xiang Li, and Quanzheng Li. Surf2ct: Cascaded 3d flow matching mod- els for torso 3d ct synthesis from skin surface. arXiv preprint arXiv:2505.22511, 2025. [ Yoon et al., 2025b ] Siyeop Yoon, Sifan Song, Pengfei Jin, Matthew Tivnan, Yujin Oh, Sekeun Kim, Dufan Wu, Xi- ang Li, and Quanzheng Li. Cascaded 3d diffusion models for whole-body 3d 18-f fdg pet/ct synthesis from demo- graphics. In International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 99â109. Springer, 2025. [ Yu et al., 2024 ] Yongrui Yu, Hanyu Chen, Zitian Zhang, Qiong Xiao, Wenhui Lei, Linrui Dai, Yu Fu, Hui Tan, Guan Wang, Peng Gao, et al. Ct synthesis with conditional diffusion models for abdominal lymph node segmentation. arXiv preprint arXiv:2403.17770, 2024. [ Zhao et al., 2025 ] Can Zhao, Pengfei Guo, Dong Yang, Yucheng Tang, Yufan He, Benjamin Simon, Mason Belue, Stephanie Harmon, Baris Turkbey, and Daguang Xu. Maisi-v2: Accelerated 3d high-resolution medical im- age synthesis with rectified flow and region-specific con- trastive loss. arXiv preprint arXiv:2508.05772, 2025. [ Zheng et al., 2025 ] Jiaxu Zheng, Meiman He, Xuhui Tang, Xiong Wang, Tuoyu Cao, Tianyi Zeng, Lichi Zhang, and Chenyu You. 3d wavelet latent diffusion model for whole-body mr-to-ct modality translation. arXiv preprint arXiv:2507.11557, 2025. [ Zhou et al., 2025 ] Xuanru Zhou, Cheng Li, Shuqiang Wang, Ye Li, Tao Tan, Hairong Zheng, and Shanshan Wang. Generative artificial intelligence in medical imag- ing: Foundations, progress, and clinical translation. arXiv preprint arXiv:2508.09177, 2025. [ Zhu et al., 2024 ] Lingting Zhu, Noel Codella, Dongdong Chen, Zhenchao Jin, Lu Yuan, and Lequan Yu. Genera- tive enhancement for 3d medical images. arXiv preprint arXiv:2403.12852, 2024.