CT-GRAPH: Graph-Based CT Modeling

• CT-GRAPH is a framework that represents CT data as graphs, where nodes denote image slices, patches, or anatomical entities and edges encode clinically meaningful relations.
• It supports diverse applications including lung cancer risk prediction, low-dose reconstruction, artifact reduction, and report generation by capturing both spatial and spectral correlations.
• Graph-based formulations in CT improve diagnostic accuracy and computational efficiency by integrating structured graph models with conventional imaging techniques.

Searching arXiv for recent and canonical uses of “CT-GRAPH” and closely related graph-based CT papers. CT-GRAPH denotes a class of graph-based formulations in which computed tomography data, CT-derived features, or CT-associated semantics are represented as graphs whose nodes and edges encode task-specific relations. In the literature considered here, the term is used for slice-graph modeling of pulmonary nodules, graph processing on sinograms, manifold and Laplacian regularization for low-dose reconstruction, hierarchical anatomy graphs for report generation, and several related constructions for lesion grounding, metal artifact reduction, spectral fusion, and segmentation. Taken together, these works suggest that CT-GRAPH is best understood as a family of graph abstractions for CT rather than a single standardized architecture (Fu et al., 2023, Mahmood et al., 2016, Kalisch et al., 7 Aug 2025).

1. Scope of the term and recurrent graph construction patterns

Across the cited CT literature, CT-GRAPH consistently follows one design principle: a CT problem is factorized into entities that become graph nodes, and clinically or physically meaningful relations become edges. The entities vary substantially by task. In some papers, nodes are axial slices of a lung nodule; in others they are overlapping sinogram patches, image patches on a manifold, anatomical regions, lesion descriptions, metal-boundary pixels, Gaussian primitives, or extrema in a 2D histogram (Fu et al., 2023, Xia et al., 2020, Li et al., 17 May 2026).

Representative formulation	Graph object	CT task
AE-GCN (Fu et al., 2023)	Nodule slices with intra-nodule edges	High-risk factor prediction in stage I NSCLC
Sinogram denoising (Mahmood et al., 2016)	Overlapping sinogram patches	Low-dose / limited-data reconstruction
MAGIC (Xia et al., 2020)	Patch-manifold graph	LDCT reconstruction
CT-GRAPH report generation (Kalisch et al., 7 Aug 2025)	Fine, coarse, and global anatomical hierarchy	Chest CT report generation
GraphMAR (Li et al., 17 May 2026)	Geometric density graph and polar artifact graph	Image-domain metal artifact reduction
Extremum graph fusion (Sharma et al., 20 Aug 2025)	Maxima–saddle graph in histogram space	Spectral PCCT volume fusion

This multiplicity is central to the term’s meaning. CT-GRAPH does not identify a single graph neural network layer, a single topology, or a single modality stage. It instead denotes a modeling strategy in which graph structure is introduced at whatever level is most faithful to the CT task: acquisition space, image space, feature space, anatomical space, or attribute space. A frequent misconception is that CT-GRAPH implies only slice-based GCN classification; the cited record does not support that restriction.

A separate paper uses “CT-graph” to mean the “configurable tree graph” for lifelong reinforcement learning. That usage concerns partially observable and distal reward environments and is unrelated to computed tomography (Soltoggio et al., 2023).

2. Slice-graph formulations for volumetric CT classification

A canonical imaging use of CT-GRAPH appears in lung cancer risk prediction. In “High-risk Factor Prediction in Lung Cancer Using Thin CT Scans: An Attention-Enhanced Graph Convolutional Network Approach” (Fu et al., 2023), the task is patient-level prediction of pathological high-risk factors for stage I non–small cell lung cancer from preoperative thin-slice CT. Each pulmonary nodule is labeled $Y_k \in \{0,1\}$, where $Y_k=1$ denotes pathological high-risk factors including micropapillary component, solid predominant pattern, complex glandular patterns, and vascular tumor thrombus. The study uses 483 nodules from 426 patients.

The graph construction is explicit. For a nodule $N_k$, each CT slice containing the lesion is cropped to a $60\times 60$ patch around the annotated center, and each slice patch $c_k^i$ becomes a node. Node features are extracted by a pre-trained VGG backbone with a Convolutional Block Attention Module inserted in the second block. The resulting slice feature is

$$x_k^i = \text{SecondLastLayer}\big(V(c_k^i)\big) \in \mathbb{R}^{1\times 512}.$$

Intra-nodule edges are then defined by one of three connectivity strategies: star graph, chain graph, or fully connected graph. The GCN stage is a two-layer spectral model with normalized adjacency $\hat A$,

$$H^{(l+1)} = \sigma\big(\hat A H^{(l)} W^{(l)}\big),$$

with a $512 \rightarrow 32$ first layer and a $32 \rightarrow 2$ second layer, followed by slice-level probabilities averaged to a nodule-level prediction,

$Y_k=1$0

Attention is confined to the CNN feature extractor; the paper explicitly notes that there is no node-level attention inside the GCN (Fu et al., 2023).

The best configuration used a fully connected intra-nodule graph and all slices. On the test set, AE-GCN achieved AUC 0.9524, Accuracy 0.9355, Sensitivity 0.8462, Specificity 0.9104, PPV 0.7857, NPV 0.9385, and F1 0.8148. Attention and graph design were both consequential: channel-only attention gave AUC 0.8123, spatial-only 0.8352, no attention 0.7732, and both together 0.9524; graph variants ranged from AUC 0.9207 for star/5 slices to 0.9524 for fully connected/all slices (Fu et al., 2023).

Model	AUC	Accuracy
AE-GCN	0.9524	0.9355
2D ResNet18	0.7738	0.7634
2D DenseNet	0.7308	0.6989
3D ResNet18	0.7692	0.7957
ViT	0.8186	0.7312
ViT+GCN	0.8582	0.7849

A closely related chest CT classifier, CT-SSG, extends the same representational idea to 3D multi-label abnormality classification by treating a volume as a structured graph of 80 non-overlapping axial slice triplets. Node features are extracted by an ImageNet-pretrained ResNet-18, edges connect nodes with $Y_k=1$1, and spectral graph convolution uses Chebyshev filters. The best reported setting used sparse topology with $Y_k=1$2, filter size $Y_k=1$3, and one spectral block, reaching macro F1 $Y_k=1$4 on CT-RATE, with reported cross-dataset results of F1 $Y_k=1$5 on Rad-ChestCT and $Y_k=1$6 on CT-HCL (Piazza et al., 12 Oct 2025). This suggests that slice-graph CT-GRAPH formulations are not limited to lesion-centric graphs; they also support scan-level reasoning over long-range inter-slice dependence.

3. Reconstruction-space CT-GRAPH: sinograms, patch manifolds, and Laplacians

A second major lineage places the graph directly in the reconstruction pipeline. In “Graph Based Sinogram Denoising for Tomographic Reconstructions” (Mahmood et al., 2016), the graph is built on the sinogram itself. For a sinogram $Y_k=1$7, each pixel-centered $Y_k=1$8 overlapping patch becomes a node, giving $Y_k=1$9 vertices; in the reported experiments, $N_k$0 had size $N_k$1, so the graph had 3420 nodes. Each patch connects to its $N_k$2 nearest neighbors by Euclidean distance, with edge weights

$N_k$3

Denoising is posed as graph total variation: $N_k$4 This graph denoised sinogram is then passed to FBP, ART, or SIRT. The reported numerical study states that graph denoised sinogram always minimizes the error measure and improves the accuracy of the solution compared with regular reconstructions, with especially large gains for smooth phantoms; for example, for the smooth phantom at RN = 0.08, FBP error decreased from 13.16 to 4.83 and SIRT error from 6.65 to 3.82 (Mahmood et al., 2016).

In “MAGIC: Manifold and Graph Integrative Convolutional Network for Low-Dose CT Reconstruction” (Xia et al., 2020), the graph is moved from the sinogram to the image patch manifold. Overlapping $N_k$5 image patches become nodes, each connected to 8 nearest neighbors with Gaussian weights. The unrolled update combines a physics-based data term, a spatial CNN branch, and a graph-convolution branch on the patch graph: $N_k$6 The graph branch uses normalized adjacency $N_k$7 and two GCN layers. On the NIH-AAPM-Mayo Clinic Low-Dose CT Grand Challenge data, MAGIC reported $N_k$8 dB and $N_k$9 SSIM at 10% dose, compared with LEARN at $60\times 60$0 dB and $60\times 60$1. Its semi-supervised version, MAGIC-Semi, retained strong performance with only 10% labeled slices, using a projection loss for unlabeled data (Xia et al., 2020).

A more explicitly resource-constrained variant appears in “Parameter-Efficient CT Reconstruction via Deep Graph Laplacian Regularization” (Radhakrishnan et al., 25 May 2026). Here each image pixel is a node in an 8-connected graph, and the regularizer is quadratic: $60\times 60$2 with learned feature-based weights

$60\times 60$3

Embedded in a Proximal Forward-Backward Splitting framework with three lightweight CNN modules, Deep GLR reported 30.70 dB PSNR on LoDoPaB-CT, a 6.33 dB improvement over filtered backprojection, using 91,848 parameters trained on 1000 samples. The paper reports 5.8 times better parameter efficiency and 30 times better data efficiency per dB improvement relative to benchmark methods, while also noting that a 13 dB gap remains versus state-of-the-art models (Radhakrishnan et al., 25 May 2026).

These three formulations share a common pattern: graph structure functions as a prior on nonlocal consistency, but the locus of that prior shifts. It can regularize projection data before reconstruction, act on image-patch manifolds during unrolled reconstruction, or regularize image grids directly through a learned Laplacian.

4. Geometry-aware and primitive-based graphs for artifacts and sparse views

In metal artifact reduction, CT-GRAPH is used to encode explicit acquisition geometry within the image domain. “GraphMAR: Geometry-Aware Graph Learning Framework for Spatially Adaptive CT Metal Artifact Reduction” (Li et al., 17 May 2026) begins from a metal mask obtained by thresholding at 2800 HU. Boundary pixels from different metal components form a geometric graph; line segments between inter-implant boundary pairs are accumulated into a geometric density graph $60\times 60$4, which coarsely localizes artifact-prone regions. At feature level, GraphMoE constructs a polar-coordinate artifact graph whose nodes are feature-map locations and whose edges encode similarity in angular direction and radial distance relative to the nearest metal pixel. The routing map learned on this graph is aligned to the geometric density prior through

$60\times 60$5

On simulated dental CT, GraphMAR$60\times 60$6 improved ProCT from 40.63/98.21 to 40.98/98.47 in average PSNR/SSIM, and on real dental CT it reduced FID from 115.83 to 108.73. The paper also reports that $60\times 60$7 experts was optimal and that the additional inference time per $60\times 60$8 image was $60\times 60$9 ms on RTX 4090 (Li et al., 17 May 2026).

Sparse-view CT has motivated a different graph construction on 3D Gaussian primitives. In “GR-Gaussian: Graph-Based Radiative Gaussian Splatting for Sparse-View CT Reconstruction” (Yuluo et al., 4 Aug 2025), each Gaussian center is a node, KNN relations define the graph, and local density smoothness is enforced by

$c_k^i$0

The method introduces a Denoised Point Cloud Initialization Strategy and a Pixel-Graph-Aware Gradient Strategy, the latter augmenting the splitting criterion with density differences $c_k^i$1 over neighboring Gaussians. With $c_k^i$2 Gaussians and $c_k^i$3 neighbors, the reported gains over R$c_k^i$4-Gaussian are 0.67 dB PSNR and 0.011 SSIM on X-3D, and 0.92 dB PSNR and 0.021 SSIM on a real-world dataset (Yuluo et al., 4 Aug 2025).

Both methods use graphs to reintroduce physically meaningful long-range structure that image-domain convolutions or view-wise gradients alone do not expose. One does so through metal geometry and polar artifact organization; the other through neighborhood relations among radiative primitives.

5. Anatomical, semantic, and segmentation graphs

Graph formulations are also used to organize CT knowledge rather than only CT intensities. In “CT-GRAPH: Hierarchical Graph Attention Network for Anatomy-Guided CT Report Generation” (Kalisch et al., 7 Aug 2025), a chest CT scan is mapped to a hierarchy of fine anatomical nodes, coarse system nodes, and a global node. Fine and coarse features are obtained by mask-guided pooling from frozen pretrained 3D encoders using TotalSegmentator masks; graph attention then aggregates fine-level information into coarse nodes and coarse information into the global node before conditioning LLaMA2-7B. On CT-RATE, the method reported BLEU-4 0.2467, ROUGE-L 0.3126, METEOR 0.4205, CE Precision 0.396, CE Recall 0.248, and CE F1 0.296, compared with the best baseline CE F1 of 0.217. An ablation reported CE F1 0.216 for a random graph, 0.253 for a single-level hierarchy graph, and 0.296 for the full hierarchical graph (Kalisch et al., 7 Aug 2025).

“GLeVE: Graph-Guided Lesion Grounding with Proposal Verification in 3D CT” (Jiang et al., 21 May 2026) pushes the graph into the report itself. Each lesion description parsed from the radiology report becomes a semantic graph containing lesion, organ, sub-location, and attribute nodes, with relation-aware attention

$c_k^i$5

The resulting lesion embedding is converted into lesion-wise queries, which are matched to anatomy-aware CT proposals and then refined by an octree autoregressive decoder. On AbdomenAtlas 3.0 with full supervision, GLeVE reported Dice 65.8%, HD95 69.5 mm, Lesion Recall 76.2%, and Lesion Localization Score 38.9%. Ablation showed that removing lesion semantic graph reasoning (“w/o LeQu”) reduced Dice from 65.8% to 59.5% and LLS from 38.9% to 32.8% (Jiang et al., 21 May 2026).

A third graph-centered anatomical application is segmentation. “Deep LOGISMOS: Deep Learning Graph-based 3D Segmentation of Pancreatic Tumors on CT scans” (Guo et al., 2018) uses a contextual UNet on three adjacent 2D patches, refines the output with GMM and morphology, and then builds a LOGISMOS graph around the refined boundary. Node costs are derived directly from UNet probabilities along graph columns: $c_k^i$6 A max-flow algorithm yields the globally optimal surface. On 51 CT scans, with 30 used for training and 21 for testing, Deep LOGISMOS reached DSC $c_k^i$7 and RVD $c_k^i$8, significantly improving over contextual UNet and LOGISMOS alone (Guo et al., 2018).

These examples show that CT-GRAPH can encode anatomy at several levels: as an organ hierarchy for language generation, as a lesion relation graph for report-to-volume grounding, or as a surface graph for globally optimal segmentation.

6. Topological, spectral, and compressive interpretations

Some CT-GRAPH formulations operate neither on voxels nor on learned semantic tokens, but on derived topological or transform domains. “Topology-Aware Volume Fusion for Spectral Computed Tomography via Histograms and Extremum Graph” (Sharma et al., 20 Aug 2025) constructs a 2D histogram from a pair of photon-counting CT volumes, computes a Morse–Smale complex on the histogram density, and extracts a maximum extremum graph whose nodes are maxima and saddles and whose edges are ascending separatrices. The extremum graph is reduced by a minimum spanning tree and longest-path extraction, then converted into a 1D parameterization used to fuse multiple spectral channels into a single representative volume. For the human heart example, neighboring energy levels had correlations 0.98–0.997, whereas the selected pair 40 keV and 110 keV had correlation 0.692 and a richer histogram structure (Sharma et al., 20 Aug 2025).

“Graph-based compensated wavelet lifting for 3-D+t medical CT data” (Lanz et al., 2023) uses graph operators for scalable representation of time-resolved cardiac CT. A bipartite graph connects voxels in adjacent frames or slices, with prediction and update matrices derived from similarity-weighted random walks. On cardiac CT data with 10 time steps and 127 slices per volume, the proposed graph-based compensation outperformed a traditional mesh-based approach of motion compensation by approximately 11 dB in terms of PSNR of the lowpass band, and the mean energy of the highpass band decreased by around 30% (Lanz et al., 2023).

Taken together, these works indicate that CT-GRAPH can serve not only discriminative modeling and reconstruction, but also topology-aware fusion and scalable compression. A plausible implication is that graph structure in CT is best regarded as a representational device whose utility depends on where redundancy, adjacency, or hierarchy naturally resides: in slices, patches, projections, organs, reports, primitives, or histogram extrema.