PaperlayerView exact revision on arXiv

Source-linked AI summary

Agentic AI Reasoning and Learned Sampling for Task-Aware Projection Selection in Sparse-View CT: Predicting Hemodynamics

Mahmoud Elhadidy, Roshan M. D'Souza, Amirhossein Arzani

arXiv:2610.39003v1cs.AIphysics.flu-dyn

TL;DR

Conventional sparse-view CT sampling prioritizes image reconstruction, while task-specific view selection for hemodynamic prediction remains an open question. This study compares fixed, learned, and agentic view-selection approaches under view budgets, finding that task-aware masks improve TAWSS prediction most under extreme sparsity. The findings also show that inference can remain accurate with views that produce poor image reconstructions.

  • Problem

    Conventional sparse-view CT patterns prioritize image reconstruction, and prior task-specific mask work focused on classification rather than hemodynamic prediction.

    Source: §1 · p. 2

  • Method

    The study compares fixed sampling with learned and agentic view selectors that choose views under the same budgets for direct hemodynamic prediction.

    Source: §1 · p. 3

  • Results

    Agentic AI and Soft topK reduced single-view spatially resolved TAWSS error approximately 68–69% compared with fixed sampling, while task-aware selection also benefited spatially averaged TAWSS at very small view budgets.

    Source: §5 · p. 31

  • Takeaways & Limitations

    Within the simulated setting, TAWSS can be predicted from one or two carefully selected views even when those views produce poor image reconstructions.

    Source: §4 · p. 29

    The study uses idealized synthetic data and one fan-beam acquisition geometry; validation on patient-specific and noisy clinical data remains needed.

    Source: §4 · p. 30, §5 · p. 31

Abstract

from arXiv · show

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.

1 Introduction

Sparse-view CT reduces projection counts to limit radiation dose, but conventional view patterns prioritize image reconstruction rather than downstream hemodynamic prediction. This study develops task-aware view selection for hemodynamic targets and compares learned selectors with an agentic AI selector.

1 Introduction

  • Views best for image reconstruction are not necessarily optimal for downstream prediction, motivating task-aware sparse-view acquisition.Fixed uniform or golden-angle patterns are common, while sparse-view CT reduces projections and can introduce imaging artifacts.Source: §1 · p. 2
  • Prior task-specific mask studies focused on classification, leaving hemodynamic prediction as an unaddressed, more complex task.The paper frames this gap around predicting hemodynamics such as velocity and wall shear stress from CTA or concentration data.Source: §1 · p. 2
  • Task-aware selection improves hemodynamic prediction over fixed baselines, with gains increasing as view counts decrease.For spatially resolved TAWSS, the gain is largest at K = 1; spatially averaged TAWSS benefits at very small view budgets.Source: §1 · p. 3
  • The framework selects views under a fixed budget and compares an agentic selector with learned and fixed methods under the same budgets.The agent proposes, evaluates, and refines candidate view sets; the study also analyzes its reasoning and mask evaluations.Source: §1 · p. 3
  • The study evaluates both spatially resolved TAWSS fields and spatially averaged TAWSS as downstream objectives.The acquisition is designed around a downstream physical quantity rather than image quality, and the selection mechanisms may apply to other projection-derived tasks.Source: §1 · p. 3

2.1 Problem Formulation

The formulation maps sparse fan-beam measurements directly to hemodynamic targets without reconstructing an image. It aggregates measurements across time and selected central slices, then jointly chooses a fixed-budget view mask and trains the predictor for the selected target.

2.1 Problem Formulation

  • The concentration volume is a voxelized snapshot at each of Nt time instants, and each slice is modeled with 2D fan-beam projections over candidate angles.The projection operator acts on a slice in the x–z plane; Nv is the number of candidate views and Nd is the number of detector elements.Source: §2.1 · p. 3
  • The two targets are a spatially resolved TAWSS field over the vessel wall and a single spatially averaged TAWSS scalar per case.TAWSS is averaged over the cardiac cycle.Source: §2.1 · pp. 3–4
  • The input is aggregated across cardiac-cycle frames and central slices containing contrast, yielding one network measurement per case.Because projection is linear, this equals the fan-beam sinogram of the frame- and slice-averaged concentration distribution.Source: §2.1 · p. 4
  • Sparse acquisition retains K distinct views from the candidate set, represented by a binary mask applied to the sinogram.The same view selection applies across slices and time frames; the per-slice, per-frame sinograms remain available for image reconstruction.Source: §2.1 · p. 5
  • The predictor directly maps the masked sinogram to the hemodynamic target, with the view mask and predictor parameters jointly optimized for the task.The mask is optimized for hemodynamic prediction accuracy rather than image reconstruction fidelity, and each task is optimized separately.Source: §2.1 · p. 5

2.2 Dataset and CTA Simulation

The study simulates aneurysm contrast transport and generates synthetic fan-beam CTA projections, then aggregates these measurements across time frames and central slices for prediction.

2.2 Dataset and CTA Simulation

  • 220 CFD simulations varied aneurysm geometry, inflow conditions, and Reynolds numbers of 260–570; five time frames per case modeled contrast transport and washout.The five frames were sampled at quarter-cycle spacing and aggregated into the network input.Source: §2.2 · p. 5
  • Each case was resampled onto a 64 × 64 × 64 Cartesian grid and modeled as 64 independent 2D fan-beam axial scans.Each slice’s rays remain within its plane, so its sinogram receives no information from neighboring slices.Source: §2.2 · p. 5
  • The fan-beam forward model integrates contrast concentration along each source-to-detector ray and produces detector readings for each view.The model used ASTRA’s fanflat geometry, with source-to-isocenter and source-to-detector distances of 600 and 1100 mm.Source: §2.2 · p. 6
  • The acquisition used 180 views at 2° increments and 64 detector elements per view, with a monochromatic, noise-free forward model.Beam hardening, scatter, and detector blur were omitted; measurement noise was added separately.Source: §2.2 · p. 6
  • Network inputs average five cardiac-cycle frames over 32 central slices containing contrast, and the same selected views apply to the whole slice stack.Slices outside this central region are almost entirely zero; TAWSS targets are represented on a 64 × 64 planar grid.Source: §2.2 · p. 6

2.3 FNO Architecture for Wall Shear Stress Prediction

The paper uses a two-dimensional Fourier neural operator to map masked sinograms to either a spatial TAWSS field or a scalar spatial average.

2.3 FNO Architecture for Wall Shear Stress Prediction

  • The FNO is used for operator learning, mapping sinogram measurements directly to the hemodynamic targets.The architecture is motivated by function-space mappings and prior applications to hemodynamic inference.Source: §2.3 · p. 6
  • The two-dimensional FNO maps masked sinograms to a 64 × 64 spatially resolved TAWSS field or, after feature pooling, a single scalar.The network uses four Fourier layers, 32 hidden channels, and 16 retained Fourier modes per dimension.Source: §2.3 · p. 6
  • A separate predictor is trained for each target using mean squared error loss.The loss compares predicted and reference outputs, with 64 × 64 outputs for the field target and one for the scalar target.Source: §2.3 · p. 7

2.4 View Selection

The section compares two fixed view-sampling baselines with three task-aware selection methods. Learned methods select views through differentiable gates or an agent’s iterative search over discrete subsets.

2.4 View Selection

  • The five methods select K views from candidate views: uniform and golden-angle sampling are fixed, while Gumbel, Soft top-K, and Agentic are learned or task-aware.Figure 2 presents each method’s selection mechanism, resulting views, and an FNO predictor.Source: Figure 2 · p. 7, §2.4 · p. 8

2.4.1 Gumbel–Softmax Selection (Gumbel)

  • Gumbel–Softmax uses K trainable selection units with relaxed view weights, enabling joint backpropagation with the predictor before pruning to K distinct views.If preferred views duplicate, the method removes duplicates and fills the remaining budget using target-specific view rankings.Source: §2.4.1 · pp. 8–9

2.4.2 Soft Top-K Selection

  • Soft top-K trains continuous sigmoid gates through a soft branch and a hard K-view branch, then freezes the top-K subset and fine-tunes the predictor.A straight-through estimator passes gradients through the hard branch; the soft branch is discarded after freezing.Source: §2.4.2 · pp. 9–10

2.4.3 Agentic AI Selection

  • Agentic AI proposes discrete subsets, trains and validates a predictor for each, then uses validation errors and search history to guide further proposals.The agent selects a final promising subset when it judges further exploration unlikely to help; the predictor is then trained with that fixed subset.Source: §2.4.3 · pp. 10–11

2.5 Training Protocol and Hyperparameters

All methods use the same FNO predictor and training schedule, with learned selectors jointly optimized only during an initial phase.

2.5 Training Protocol and Hyperparameters

  • All methods use the same four-layer FNO with 32 hidden channels and 16 × 16 retained Fourier modes.The dataset contains 200 training cases and 20 test cases; Agentic uses a validation subset drawn from training.Source: §2.5 · p. 11
  • Models train for 3000 epochs with AdamW, a 3 × 10^-3 learning rate, 10^-4 weight decay, batch size 16, cosine annealing, and MSE loss.Source: §2.5 · p. 11
  • Soft top-K and Gumbel jointly optimize selector and predictor for 500 epochs, then freeze exact-K masks and train predictors alone for 2500 epochs.Uniform and golden-angle masks remain fixed throughout training.Source: §2.5 · p. 11

3 Results

Across view budgets, task-aware selection most benefits spatially resolved TAWSS under extreme sparsity, while spatially averaged TAWSS benefits from optimized selection mainly at the smallest budgets. Results report mean relative L2 error over five runs for each method, budget, and target.

3 Results

  • Budgets were K ∈ {90, 45, 15, 2, 1}; uniform and golden-angle sampling use the same view at K = 1, and results summarize five runs.The five agentic runs are independent searches, each starting with a fresh agent context.Source: §3 · p. 11

3.1 Spatially Resolved TAWSS

  • For spatially resolved TAWSS, Agentic and Soft top-K have the lowest mean error at every budget, while methods lie within one standard deviation at K = 90.At K = 15, errors range from 0.0384 (Agentic) to 0.0439 (Golden), and no method's error rises by more than about 21% relative to K = 90.Source: §3.1 · pp. 11, 13, Figure 3 · p. 12, Table 1 · p. 13
  • At K = 1, Agentic and Soft top-K reduce spatially resolved TAWSS error by about threefold versus the other methods; at K = 2, their errors are 0.0544 and 0.0554.Agentic and Soft top-K errors increase only 1.7× and 1.8× from K = 90 to K = 1, versus 4.8×, 4.9×, and 5.0× for Gumbel, Uniform, and Golden.Source: Figure 3 · p. 12, §3.1 · p. 13, Table 1 · p. 13
  • At K = 2 and K = 1, the three highest-error methods produce visibly noisier spatially resolved TAWSS maps, especially in one test case.The qualitative examples are two randomly selected test cases; their per-case errors differ from the run-averaged Table 1 results.Source: §3.1 · p. 13, Figure 4 · p. 14

3.2 Spatially Averaged TAWSS

  • At K ≥15, view selection gives no consistent improvement for spatially averaged TAWSS; at K ≤2, Agentic, Gumbel, and Soft top-K show distinct improvements.Agentic is the only method that never ranks below third across budgets.Source: §3.2 · pp. 13, 15, Figure 5 · p. 15
  • From K = 90 to K = 1, spatially averaged TAWSS error grows 1.7× for Soft top-K and about 2.6× for Agentic and Gumbel, versus 5.8× for Uniform and 9.3× for Golden.There is no clear winner across all budgets.Source: §3.2 · p. 13, Figure 5 · p. 15

3.3 Interpretation of Selected Views

Learned masks tend to favor concentration-transition regions, but distinct angular patterns can achieve similar accuracy. Across targets, masks share some heavily used regions, while overlap between budgets drops sharply at the sparsest levels.

3.3 Interpretation of Selected Views

  • At K = 45, learned methods tend to select views between high- and low-concentration regions, where gradients are steepest and aneurysm cases differ most.The selected views are shown over the concentration volume; each ray marks a projection angle.Source: §3.3 · p. 15, Figure 6 · p. 16
  • At K = 45, Gumbel forms two view clusters, Soft top-K forms three or four contiguous clusters, and Agentic often uses mirrored or symmetric arrangements.Agentic is more consistent across seeds than Soft top-K, while Gumbel is most consistent but least accurate among the three learned methods.Source: Figure 6 · p. 16, §3.3 · pp. 17–18
  • At K = 15, Gumbel concentrates views in one region, Soft top-K retains three or four groups, and Agentic retains symmetric arrangements with the best accuracy at this budget.Agentic becomes less consistent than at K = 45.Source: Figure 7 · p. 17, §3.3 · p. 18
  • At K = 2, some seeds place both views adjacent to each other; very different masks can reach similar accuracy, so no single best set is evident.The most repeatable method at this budget is also the least accurate.Source: §3.3 · p. 18, Figure 8 · p. 18

3.4 Agreement Between Masks Across Tasks and Budgets

  • Both targets heavily use 80°–160° and least use regions around 0°–20°, 170°–190°, and 350°–360°; overlap between budgets falls to 0.15–0.35 from K = 2 to K = 15.Agentic's overlap in that sparsest budget transition is almost double that of the other methods.Source: §3.4 · p. 19, Figure 9 · p. 19

3.5 Agentic AI Behavior and Efficiency

The agentic search explored and refined discrete view masks through sequential trials, with recurring reasoning patterns across runs and budgets. Searches typically improved through multiple candidates, and selected masks could be reproduced by retraining logged trials.

3.5 Agentic AI Behavior and Efficiency

  • The search analysis tracked 390 trials across 50 runs, coding candidate constructions into 17 reasoning categories that could overlap.The runs were independent by task and budget, and the logs recorded masks, validation errors, and stated reasoning.Source: §3.5 · p. 20, §3.5.1 · p. 20, Table 3 · p. 22

3.5.1 Reasoning Concepts and Sequential Reasoning

  • Across searches, uniform placement commonly opened the run but rarely won, while block placement was the most frequently selected final concept and orthogonal placement was never selected.Uniform/grid placement appeared in 58 trials and opened 60% of runs but produced the best validation error once; block/contiguous placement was selected 10 times among 57 proposals.Source: §3.5.1 · pp. 20, 24, Figure 11 · p. 23
  • Sensitivity or robustness screening had the highest selection-to-proposal ratio, and was commonly followed by local refinement.Screening was selected in 6 of 26 trials; local refinement followed it in 50% of switches.Source: §3.5.1 · pp. 20, 24, Figure 11 · p. 23
  • Exploit, explore, and wildcard trials made up roughly 55%, 30%, and 15% of proposals, while 41% declared only earlier trials as their basis.Another 31% declared both prior trials and a separate analysis, 18% analyses alone, and 9% no basis.Source: Figure 10 · p. 21, §3.5.1 · p. 24

3.5.2 Agent Search Trajectories

  • Search trials evaluated discrete candidate masks, so validation errors could rise between trials even as the cumulative best error improved.The cumulative-minimum curves track the lowest error encountered up to each trial.Source: §3.5.2 · p. 24, Figure 12 · p. 25
  • Spatially resolved searches averaged 8.36 trials and spatially averaged searches 7.24; the selected candidate was typically found before termination, with several new best candidates during search.New best errors occurred an average of 3.04 times after the initial trial for spatially resolved TAWSS and 2.48 times for spatially averaged TAWSS.Source: §3.5.2 · pp. 24–25, Figure 12 · p. 25

3.5.3 Agentic AI Reproducibility: Human Validation

  • Manual reruns reproduced logged validation Rel-L2 errors to within 0.02% for a two-view field trial and 3.95% for a 15-view scalar trial.Both trials were retrained from scratch using the agent's recorded settings.Source: §3.5.3 · p. 26

3.6 Computational Cost

Task-aware masks could preserve hemodynamic prediction performance while producing poorer image reconstructions, and all methods became vulnerable as measurement noise increased. The agentic selector also required additional compute for its candidate search.

3.6 Computational Cost

  • At K = 45, Agentic selection took about 24 minutes, 64% longer than Uniform, because it evaluated candidate subsets before final predictor training.Each candidate used a short 200-epoch run of about one minute; final predictor training cost the same across methods.Source: §3.6 · p. 26

3.7 Robustness to Measurement Noise

  • All methods degraded quickly as sinogram noise increased; at the highest noise level, and already at moderate noise for K = 2, predictions collapsed toward near-constant outputs.At low noise, task-aware methods generally remained competitive, while uniform was worst especially at K = 2.Source: §3.7 · pp. 26–27, Figure 13 · p. 27

3.8 CTA Image Reconstruction

  • At K = 90, uniform sampling gave the highest reported reconstruction quality, but task-aware masks could outperform it at sparse budgets and did not necessarily preserve image quality.For the first slice at K = 90, uniform reached 53.6 dB PSNR versus 45.8 dB for golden-angle; at K = 2, Gumbel reached 17.6 versus uniform's 13.9 dB for the second slice.Source: §3.8 · pp. 27, 29, Figure 14 · p. 28
  • Below about 21 dB PSNR, reconstructions from every method no longer showed a recognizable concentration field, and the Agentic reconstruction was strongly distorted by K = 45.The comparison used FBP reconstructions against full 180-view references, with one representative mask per method.Source: §3.8 · pp. 27, 29, Figure 14 · p. 28

4 Discussion

The discussion argues that task-aware selection can improve hemodynamic prediction under sparse acquisition, while the most useful masks and gains depend on the target and view budget. The study also identifies limits in simulated data, measurement noise, acquisition geometry, and reconstruction quality.

4 Discussion

  • Task-aware learned and agentic masks improved TAWSS prediction over fixed baselines, with gains increasing as views decreased for spatially resolved TAWSS, while scalar-target gains were concentrated at K = 1.For spatially averaged TAWSS, fixed baselines were as accurate or slightly better at K ≥15.Source: §4 · p. 29
  • The selected views clustered near sharp concentration changes between the aneurysm jet and surrounding low-concentration region, but degraded reconstructed-image quality.The discussion notes possible practical implementation through angular tube-current modulation, contiguous projection blocks, or adaptive collimation.Source: §4 · pp. 29–30
  • Several distinct view subsets achieved comparable prediction accuracy, so the study does not identify a uniquely best mask.The discussion relates this non-uniqueness to prior optimal sensor-placement and sampling-mask studies.Source: §4 · p. 29
  • The agentic search adapted using prior results and combined surrogate modeling, greedy selection, angular coverage, and mask refinement; it had the lowest reported mean error at four budgets.It tied Soft top-K to four decimal places at K = 90 for spatially resolved TAWSS.Source: §4 · p. 30
  • The evaluation used idealized synthetic data and one fan-beam geometry, assumed quasi-static contrast during each gantry rotation, and found substantial sensitivity to measurement noise.The study also focused its selection objective on hemodynamic prediction rather than conventional image reconstruction quality.Source: §4 · p. 30
  • Future work could examine denoising, combined image-reconstruction and hemodynamic objectives, alternative agents, and hardware-aware search constraints.The proposed hardware restrictions include limited angular regions and feasible tube-current modulation.Source: §4 · pp. 30–31

5 Conclusion

The conclusion presents task-aware CTA view selection as a way to improve hemodynamic prediction under highly sparse acquisition, while noting that clinical validation remains needed.

5 Conclusion

  • Agentic AI and Soft topK reduced single-view spatially resolved TAWSS error by approximately 68–69% versus fixed sampling, while task-aware selection also helped spatially averaged TAWSS at very small budgets.Multiple distinct subsets achieved similar accuracy, and views optimized for hemodynamic inference were not necessarily optimal for image reconstruction.Source: §5 · p. 31

Related papers

Loading 2610.39003v1…