Agentic AI Reasoning and Learned Sampling for Task-Aware Projection Selection in Sparse-View CT: Predicting Hemodynamics
Authors: Mahmoud Elhadidy, Roshan M. D'Souza, Amirhossein Arzani
Organizations: Department of Mechanical Engineering, University of Utah, Salt Lake City, UT, USA · Scientific Computing and Imaging Institute, University of Utah, Salt Lake City, UT, USA · Department of Mechanical Engineering, University of Wisconsin–Milwaukee, Milwaukee, WI, USA · Department of Biomedical Engineering, University of Utah, Salt Lake City, UT, USA
Sparse-view computed tomography (CT) reduces the number of acquired projections, but conventional sampling strategies are designed primarily for image reconstruction rather than for downstream physical inference. Here, we investigate optimal task-aware projection selection for directly predicting hemodynamics from sparse CT angiography (CTA) measurements. Using computational fluid dynamics simulations of intracranial aneurysms, we generate synthetic fan-beam CTA projections and train a Fourier neural operator (FNO) to predict two targets: the spatially resolved time-averaged wall shear stress (TAWSS) field and the spatially averaged TAWSS. The FNO was coupled with optimal input selection for simultaneous optimal sampling and training. Our goal was to investigate if agentic AI could discover optimal sampling strategies that outperform current methods. We compared fixed uniform and golden-angle sampling with three task-aware optimal selection strategies: Gumbel-Softmax, a Soft top-K method, and an agentic AI selector. The methods were compared at different view selection budgets. A detailed analysis of the AI agent was performed to understand its reasoning and performance. For spatially resolved TAWSS, Agentic and Soft top-K achieve the lowest mean errors (Agentic performs slightly better) and substantially outperform fixed sampling under extreme sparsity. For spatially averaged TAWSS, optimized selection becomes important at extremely sparse acquisitions. Analysis of the selected masks reveals multiple distinct subsets concentrated around informative regions. Moreover, masks optimized for hemodynamic prediction yield poor image reconstructions, showing that task-optimal acquisition differs fundamentally from reconstruction-optimal acquisition. This work suggests the possibility of designing sparse acquisitions directly around downstream physical quantities.
Figures & tables
Figure 1: Overview of the study. (a) Data generation: an idealized aneurysm geometry gives a pulsatile velocity field, which drives an advection–diffusion simulation of contrast concentration. The concentration field is voxelized and fan-beam sinograms are synthetically generated. The training dataset contained distinct geometry and flow rates. (b) Prediction from a fixed sparse view set: the predictor fθ:RNd×Nv→RM maps the sparse-view sinogram directly to the projected time-averaged wall shear stress (TAWSS) on the vessel wall, where M is the number of predicted outputs as introduced in Sec. 2.3 . (c) Prediction with learned view selection: the acquired subset S of the Nv candidate views is optimized together with the predictor model fθ , guided by the task loss.
Figure 2: Five mechanisms for choosing which K views of the Nv candidates are selected. Each row gives the defining equations on the left, the resulting selection in the center, and a similar predictor fθ , implemented as an FNO, on the right. (a) Uniform sampling: K equispaced angles, fixed before training. (b) Golden-angle sampling: fixed 137.5∘ increments. (c) Gumbel–Softmax selection (Gumbel). (d) Soft top- K selection. (e) Agentic AI selection (Agentic).
Figure 3: Prediction of spatially resolved TAWSS at different view budgets K . (a) Test Rel- L2 error across five runs. Boxes span the mean ±1 standard deviation, center lines indicate the mean, and whiskers span the mean ±2 standard deviations. Stars mark the lowest mean error at each budget. (b) Gap to the best method at each budget, expressed as Cohen’s d : the difference in mean error divided by the pooled standard deviation of the two methods, d=(εˉi−εˉbest)/(si2+sbest2)/2 . The best method at each budget is marked best . (c) Mean error normalized by each method’s mean error at K=90 , showing the relative increase as the view budget decreases. (d) Methods ranked by mean error at each budget, with rank 1 indicating the lowest error. At K=90 , Agentic and Soft top- K are tied. Agentic is shown as rank 1 solely for visualization purposes.
Method
K=90
K=45
K=15
K=2
K=1
Uniform
0.0391±0.0056
0.0401±0.0052
0.0432±0.0069
0.1547±0.0217
0.1925±0.0165
Golden
0.0386±0.0054
0.0413±0.0061
0.0439±0.0064
0.0706±0.0090
0.1924±0.0167
Gumbel
0.0345±0.0033
0.0391±0.0073
0.0416±0.0058
0.0738±0.0289
0.1660±0.0435
Soft top- K
0.0343±0.0045
0.0348±0.0036
0.0407±0.0062
0.0554±0.0122
0.0623±0.0129
Agentic
0.0343±0.0052
0.0347±0.0063
0.0384±0.0051
0.0544±0.0065
0.0588±0.0043
Table 1: Field prediction errors (mean ± standard deviation across five runs). Bold indicates the lowest mean error at each view budget. At K=90 , Agentic and Soft top- K have equal mean error to four decimal places and are both marked in bold. At K=1 , Uniform and Golden use the same view, and their small differences arise from non-deterministic GPU arithmetic during training.
Figure 4: Predicted spatially resolved TAWSS for two test cases, shown for all five methods at view budgets K . Each panel gives the prediction with the Rel- L2 error of that case below it, alongside the ground truth. The cases were selected at random from the test set and the per-case errors therefore differ from the run-averaged values in Table 1 .
Figure 5: Spatially averaged TAWSS errors for view budgets K . Panels follow Fig. 3 , with means and standard deviations computed over five runs. In panel (a), the error axis is logarithmic, so lower whiskers that fall below the lowest value shown, including those where the mean minus two standard deviations is negative, are truncated at the bottom of the axis.
Method
K=90
K=45
K=15
K=2
K=1
Uniform
0.0264±0.0140
0.0177±0.0061
0.0209±0.0092
0.0985±0.0526
0.1540±0.0346
Golden
0.0166±0.0061
0.0215±0.0053
0.0218±0.0081
0.0465±0.0186
0.1545±0.0317
Gumbel
0.0210±0.0067
0.0242±0.0056
0.0294±0.0116
0.0322±0.0087
0.0545±0.0106
Soft top- K
0.0260±0.0127
0.0221±0.0056
0.0285±0.0114
0.0404±0.0134
0.0435±0.0176
Agentic
0.0173±0.0063
0.0210±0.0054
0.0249±0.0049
0.0320±0.0075
0.0446±0.0112
Table 2: Scalar (spatially averaged) prediction errors (mean ± std over 5 runs). Bold marks the best value in each column. At K=1 , Uniform and Golden use the same view, and their small differences arise from non-deterministic GPU arithmetic during training.
Figure 6: Views selected by each method for the spatially resolved target at K=45 , drawn over the contrast concentration volume at t=2 s. Each ray marks one acquired projection angle and the outline shows the full 360∘ candidate ring of Nv=180 views spaced 2∘ apart. The first five columns are the five runs, each labeled with its test Rel- L2 . The last column pools the runs and is labeled with the mean ± standard deviation over them. In this last column, the rays are colored by how many of the five runs selected each view (color scale at the right).
Figure 7: Views selected by each method for the spatially resolved target at K=15 , drawn over the contrast concentration volume at t=2 s. Layout as in Fig. 6 .
Figure 8: Views selected by each method for the spatially resolved target at K=2 , drawn over the contrast concentration volume at t=2 s. Layout as in Fig. 6 .
Figure 9: Angular structure of the learned masks for Gumbel, Soft top- K and Agentic with five runs per target and budget (60 masks per target at K≤45 ). (a) Percentage of masks that place at least one view within ±5∘ of each angle, for K≤45 . The angle axis spans the full 360∘ gantry rotation. Shaded bands mark 80∘ – 160∘ and 240∘ – 280∘ . (b) Fraction of a mask that reappears at the next higher budget, averaged over runs and both targets.
Figure 10: Reasoning of the Agentic search over all 390 scored trials from 50 runs. Design concepts are coded by keyword match on the stated candidate construction and a trial may name more than one concept. Each concept is defined in Table 3 . (a) Trials naming each concept. (b) For each concept, the ratio of the trials that were finally chosen to the trials that proposed it. (c) Share of trials by role at each budget, for the spatially resolved (R) and spatially averaged (A) targets. Roles are the agent’s own labels: exploit refines a construction that already worked in that run, explore tests an unproven construction or a deliberate control, and wildcard is an unconventional idea outside the current line of reasoning, which the protocol required each run to attempt. (d) The predecessors each trial declares: 41% build only on earlier trials in the same run, 31% on both earlier trials and separate analyses, 18% on analyses alone, and 9% declare none.
Concept
Proposed
Chosen
What it means
Uniform / grid
58
1
Views placed at equal angular spacing, optionally rotated by an offset. The standard sparse-view construction, used here mostly as a starting point.
Block / contiguous
57
10
Views drawn from one or more contiguous angular bands rather than spread over the full range, including bands that wrap past 360∘ .
Control / baseline
46
2
A candidate proposed to test a claim rather than to win, such as a deliberately poor placement expected to fail.
Ridge / linear proxy
45
4
Views ranked by a cheap supervised surrogate, typically a ridge or least-squares fit from single views to the target.
Coverage / spread
42
8
Placement driven by an explicit geometric spacing criterion such as farthest-point, stratified or low-discrepancy sampling.
Local refinement
35
7
Small perturbations of an existing mask: shifting a view by a few degrees, jittering, or swapping in an angular neighbor.
Table 3: The design concepts coded in Fig. 10 . A trial is assigned a concept when its stated candidate construction contains one of the listed keywords, so a single trial may carry several concepts. Proposed is the number of trials naming the concept and chosen the number of those that became a run’s final used mask.
Figure 11: How the agents move through design concepts, using the same keyword coding as Fig. 10 over 390 trials from 50 independent searches. Concepts are ordered by their typical position in a run, and both panels share that order. (a) Percentage of trials using each concept at each trial number, with the eighth and later trials pooled. Uniform sampling dominates the first trial, while local refinement and sensitivity checks appear only from the fourth trial onward. (b) For each concept on the left, the three concepts the agent most often chooses in the next trial, ordered from most to least common and colored by that share. A switch is a concept used in the next trial but not in the current one.
Figure 12: Search trajectories of the Agentic view selector: five independent search agents per panel, one column per view budget. (a) Validation error of the candidate proposed at each trial for spatially resolved TAWSS. (b) The lowest error reached up to that trial, i.e., the cumulative minimum of panel (a). (c) , (d) The same two views for spatially averaged TAWSS. Rings mark the trial each agent finally selected.
Method
Selection (s)
Predictor training (s)
Total (min)
Peak GPU memory (MB)
Uniform
0
892.5
14.9
561
Golden
0
892.5
14.9
561
Gumbel
In-train
895.7
14.9
562
Soft top- K
In-train
1005.3
16.8
1053
Agentic
570 †
892.5
24.4
617 ∗
Table 4: Computational cost of the view selection methods for the spatially resolved TAWSS at K=45 .
Figure 13: Robustness to additive sinogram noise for spatially resolved TAWSS. (a) K=45 views. (b) K=2 views. Bars are the mean and whiskers ±σ over five runs. The shaded region marks Rel- L2>0.5 , where the TAWSS predictor fails.
Figure 14: Filtered back-projection reconstructions from the mask each method selected for two y slices of the same case at t=2 s. Rows are the five view-selection methods and columns the view budget K . All panels are cropped to the object and the concentration is normalized to 0 – 1 within each slice. The number under each panel is the peak signal-to-noise ratio (PSNR) against that slice’s reference.
Computed tomography uses X-ray measurements from multiple angles to reconstruct cross-sectional and volumetric images for diagnosis and scientific imaging.
Sparse-view CT reduces radiation dose and scanning time by acquiring fewer projection views, but angular undersampling makes reconstruction severely ill-posed, causing streak artifacts, structural blurring, and loss of fine details. Existing supervised methods are often tied to specific sampling settings, whereas generative methods may introduce anatomically inconsistent hallucination-like structures under severe undersampling. We propose Lucid, a sparsity-adaptive, consistency-guided reconstruction framework based on a Flow Matching generative prior for sparse-view CT. Lucid is trained only on high-quality CT images to learn a continuous transport between a Gaussian distribution and the high-quality CT image distribution, independent of view sampling. During inference, the sampling sparsity level is explicitly incorporated to adapt the generative trajectory of a single pretrained model. Specifically, Lucid constructs a degradation-matched initial state by sparsity-weighted fusion of the sparse-view FBP image and Gaussian noise, performs sparsity-modulated Flow Matching updates, and applies projection-domain data-consistency correction after each prior update. Experiments under multiple sparse-view settings show that Lucid achieves stable reconstruction performance across different sampling densities, improves image quality and structural fidelity, and reduces the risk of hallucination-like structures in generative sparse-view CT reconstruction.
Jigang Duan, Jiayi Wang, Heran Wang +3
School of Mathematical Sciences, Capital Normal University, Beijing 100048, China · National Center for Applied Mathematics Beijing, Capital Normal University and Academy for multidisciplinary studies, Capital Normal University, Beijing, 100048, China
Sparse-view computed tomography (CT) reduces radiation dose and acquisition time but remains severely ill-posed because incomplete projections poorly constrain null-space information. Existing learning-based methods often estimate this information in high-dimensional image space, conflate physical measurement errors with prediction errors, and depend on fixed discretizations. We propose SONAR, a Structure-Consistent Neural Operator for Null-Space-Aware Reconstruction. Instead of recovering the full null-space component, SONAR predicts a low-dimensional null-space-aware representation from the acquired projections as pseudo-measurements. It separates measurement and pseudo-measurement residuals, lifts them into the image domain through physics operators, and applies independent neural operators to constrain their structural effects, thereby accommodating admissible errors while suppressing unsupported structures. To support cross-discretization reconstruction, an anisotropic U-shaped neural operator models the periodic angular and nonperiodic detector dimensions using direction-dependent continuous supports, while image-domain neural operators re-discretize continuous kernels on target grids. These components form an optimization-inspired unrolled network. Experiments on simulated AAPM and clinical MARS photon-counting CT data demonstrate consistent improvements across seen and unseen view settings and unseen image resolutions. On AAPM dataset, SONAR improves PSNR by 1.87dB at 62 views and by 7.63dB under zero-shot transfer to a 512×512 grid over the strongest competing methods. SONAR also achieves the best overall performance in all clinical settings evaluated, demonstrating accurate, structurally reliable, and discretization-robust sparse-view CT reconstruction.
Song Ni, Haijun Yu, Haodong Li +4
Department of Electrical and Computer Engineering, University of Massachusetts Lowell, Lowell, MA 01854
This work introduces a continuous soft near-orthogonality score and a resolution-aware saturated coverage objective for projection selection in region-of-interest focused cone-beam CT, grounded in Tuy's completeness theory. Replacing the binary hit-or-miss model of classical Tuy completeness with a graded, differentiable formulation preserves a direct link to achievable feature sizes while enabling both efficient approximate and exact optimisation. We establish that the underlying discrete decision problems are NP-complete via polynomial-time reductions from Set Cover, motivating a submodular greedy algorithm with proven (1−1/e) approximation guarantees and a mixed-integer linear program (MILP) that provides certified optimality bounds. The MILP serves as a quality certificate for the greedy solution rather than a competing optimiser. The primary empirical finding confirms this relationship: across a systematic benchmark spanning six target regions, multiple projection budgets, and four controlled occlusion conditions, the pooled median greedy-to-MILP objective ratio was 0.998, with a substantial fraction of cases certified globally optimal. A binary formulation is included as a diagnostic baseline; it strengthens hard directional completeness but is weaker on the continuous coverage scale. We additionally introduce Effective Spatial Resolution (ESR), a physically interpretable trajectory-level diagnostic that maps directional sampling gaps to achievable feature sizes. ESR correlates reliably with matched reconstruction quality across projection budgets and occlusion levels, providing a practical bridge between the selection stage and the image domain without requiring reconstruction.