1 of 18

AI-Based CT Acquisition

Parameter Prediction

A physics-simulation-driven surrogate-modelling and inverse-optimisation pipeline

L. Bhargav | 24324007 | BS-MS Physics, IIT Roorkee | PHC-351 | all results are simulated, not scanner measurements

Under Prof. Mayank Goswami

2 of 18

2

OVERVIEW

The project in one slide

Choose kV, mAs, projections and Cu filter for a new sample: image quality versus scan time.

Problem

Four acquisition parameters, three materials, many geometries; trial-and-error scanning costs time and dose.

Approach

Polychromatic Beer-Lambert + Poisson simulator with ASTRA FDK reconstruction, 3,600 runs, four surrogate regressors, constrained inverse search, re-simulation.

Headline results

GP held-out CNR R² = 0.985; Optuna mean target-CNR error 2.94% vs 16.99% for the heuristic (6 tasks, re-simulated).

Scope: simulation-only study on three materials with idealised single-slice phantoms. SNR/CNR come from the simulator; scan time and dose are model-based proxies.

AI-Based CT Acquisition Parameter Prediction | simulation-based study

3 of 18

3

PHYSICS

Polychromatic source, filtration and attenuation

The spectrum is hardened by Al and Cu filters, then attenuated energy by energy.

N₀(E) = k · mAs · [ (kV − E)/E + Σₗ wₗ G(E−Eₗ) ] · exp(−μ_Al t_Al − μ_Cu t_Cu) · ΔE

Ī(ray) = Σ_E N₀(E) · exp[ −μ_m(E) L_m − μ_inc(E) L_inc ], μ(E) = (μ/ρ)_XCOM(E) · ρ

  • k = 260 (fixed flux constant); characteristic lines only for kV > 69.5; log-log interpolation of XCOM tables
  • Matrix: PMMA 1.18, Al 2.70, Steel 7.87 g/cm³ (density scale 0.92–1.08); inclusion: Al in PMMA/Steel, PMMA in Al
  • L_m, L_inc: ray path lengths through matrix and inclusion, from ASTRA forward projection of the material masks

Why it matters

Hardening shifts mean energy upward and drives cupping; low-kV, heavily filtered beams starve Steel first.

Source: Code/ct_simulator.py; NIST XCOM (Berger et al.)

AI-Based CT Acquisition Parameter Prediction | simulation-based study

4 of 18

4

ARCHITECTURE

Pipeline from configuration to recommendation

Physics produces data; data trains surrogates; surrogates drive the search; physics checks the answer.

Phantom

+ masks

Spectrum

+ filtration

Beer-Lambert

+ Poisson

ASTRA FDK

recon

SNR, CNR,

proxies

3,600-run

dataset

XGB, RF,

MLP, GP

Grid / Optuna

search

Re-simulate

& verify

Physics simulation Machine learning Optimisation and validation

  • Inputs (8 features): material, geometry, size, density, kV, mAs/projection, N projections, Cu thickness
  • Targets (4): SNR, CNR, edge-FWHM proxy, scan-time proxy
  • Starved runs are logged and also used to train a feasibility classifier

AI-Based CT Acquisition Parameter Prediction | simulation-based study

5 of 18

5

SIMULATION

Phantom, geometry and projection generation

Material masks are forward-projected to give per-ray path lengths in each material.

  • Shapes: cylinder, ellipse (axes 1.15R, 0.85R) or block; one inclusion plus an air void
  • ROIs: background and inclusion (eroded), centre/perimeter for cupping
  • Cone-beam geometry, 1 detector row, 128×128×1 volume: a central slice only

Fig. 1 Representative reconstructed slices (project output, Figures/fig01).

AI-Based CT Acquisition Parameter Prediction | simulation-based study

6 of 18

6

SIMULATION

Photon statistics, Poisson noise and starvation

Noise is physical: counts are Poisson, so SNR scales with the square root of exposure.

N_meas ~ Poisson( Ī ), p = −ln( max(N_meas, 1) / N₀,total ), SNR ∝ √N

  • Starved if: fraction of sample rays with ≤ 2 counts > 0.25, or min expected counts < 2
  • In-distribution: 210 of 3,150 starved (6.7%); extrapolation: 53 of 450 (11.8%)
  • Extrapolation starvation: E2 low-kV/heavy-filter 32/150, E1 thick 16/150, E3 5/150
  • Why keep them: they map the feasibility boundary the optimiser must respect

Limitation

Uniform open-beam flux; no scatter, detector response or electronic noise are modelled.

Source: ct_simulator.py; Data/Raw_Sweep master CSVs

AI-Based CT Acquisition Parameter Prediction | simulation-based study

7 of 18

7

SIMULATION

Reconstruction and image-quality metrics

Verified engine: ASTRA FDK_CUDA on a single-slice cone-beam geometry for all 3,600 runs.

SNR = μ̅_bg / σ_bg CNR = |μ̅_inc − μ̅_bg| / √[(σ²_bg + σ²_inc)/2] Cupping = 100 (μ_perim − μ_ctr)/μ_perim

Scan-time proxy

T = N·(mAs/5 + 0.015) + 2 s. A model formula, not a scanner measurement.

Dose proxy

D = N · Σ E·N₀(E) · 10⁻⁶: source energy fluence, no sample attenuation. Not absorbed dose.

Edge-FWHM proxy

Width of the gradient at the inclusion edge. Not an MTF; surrogate R² only 0.73 (test).

Note: the inverse-search code uses a different approximate dose function, N·mAs·(kV/100)^2.1·exp(−0.35·d_Cu), as a small term in its objective.

Source: ct_simulator.py; main_pipeline.ipynb; recon_algorithm column of raw sweeps

AI-Based CT Acquisition Parameter Prediction | simulation-based study

8 of 18

8

DATASET

Design space, validity and splits

Latin-hypercube sampling over six continuous parameters; physically valid runs are split by material.

Set

Attempted

Valid

Starved

In-distribution

3,150

2,940

210

Train / Val / Test

​

2,058 / 441 / 441

​

Extrapolation

450

397

53

Valid per split (PMMA / Al / Steel): train 700/700/658; val 150/150/141; test 150/150/141; extrapolation 150/147/100.

  • Size 5–20 mm; density scale 0.92–1.08
  • kV 75–220; mAs/projection 0.10–1.50
  • N projections 60–360; Cu 0–2 mm
  • Extrapolation: E1 thick 20.5–25.9 mm; E2 low-kV/heavy-filter; E3 mAs 1.55–2.20, N 36–439

Source: Data/Raw_Sweep/*.csv, Training/Validation/Testing/Extrapolation CSVs (counted directly)

AI-Based CT Acquisition Parameter Prediction | simulation-based study

9 of 18

9

MACHINE LEARNING

Surrogate models and training methodology

Four regressors were compared: trees, a neural net and a Gaussian process behave differently.

Model

Key settings (as evaluated)

CV CNR R²

Train time

XGBoost

450 trees, depth 6, lr 0.05, subsample 0.85

0.970

5 s

Random Forest

400 trees, depth 18, min leaf 2

0.840

3 s

MLP 64-64-32

ReLU, Adam, early stopping, α=1e-3

0.977

9 s

Gaussian Process

ARD-RBF × const + White noise

0.979*

277 s

  • Features standardised (StandardScaler); seed 2026; train/val/test fixed before fitting
  • *GP CV: 3-fold on a 900-row subsample; other models 5-fold on the training set
  • Search-time XGBoost was refit (500 trees, depth 7, lr 0.045, colsample 1.0); its metrics are not separately tabulated
  • Starvation classifier: Random Forest (200 trees) fit on all 3,150 attempted runs, so not held out

Source: Docs/surrogate_evaluation_metrics.csv; main_pipeline.ipynb cells 4–5

AI-Based CT Acquisition Parameter Prediction | simulation-based study

10 of 18

10

RESULTS

Held-out in-distribution performance (n = 441)

The Gaussian process is the most accurate; XGBoost is the model labelled primary.

MAE

SNR

CNR

GP

0.94

1.04

MLP

1.35

1.35

XGBoost

1.30

1.59

RF

2.81

3.08

Training cost: GP 277 s vs XGBoost 5 s. Interpolation accuracy alone does not decide the choice of downstream model.

Source: Docs/surrogate_evaluation_metrics.csv (split = InDist_Test)

AI-Based CT Acquisition Parameter Prediction | simulation-based study

11 of 18

11

RESULTS

Extrapolation and failure regimes (n = 397)

XGBoost and RF degrade sharply in the unseen low-kV / heavy-filtration regime; GP and MLP hold up.

Overall CNR R²

XGB 0.907 | RF 0.803 | MLP 0.941 | GP 0.943

Reading

Interpolation scores overstate real-world robustness; check ensemble disagreement near range edges.

Source: Docs/extrapolation_by_regime.csv; surrogate_evaluation_metrics.csv

AI-Based CT Acquisition Parameter Prediction | simulation-based study

12 of 18

12

OPTIMISATION

Constrained inverse recommendation

Minimise predicted scan time subject to a CNR target, time budget and photon feasibility.

min T(mAs, N) + 25·|ĉ − 1.03c*|/c* + 0.015·D̂ s.t. T ≤ T_max, P_valid ≥ 0.80, ĉ − 0.5σ ≥ 1.01c*

  • ĉ = 0.50·XGB + 0.35·GP + 0.15·MLP (weights fixed by hand; σ = inter-model spread)
  • Search space: kV 85–218, mAs 0.18–1.48, N 75–360 (step 15), Cu 0.20–1.20 mm
  • Optuna TPE, 120 trials, seed 2026; infeasible points get graded penalties

Why re-simulate

The surrogate can be wrong exactly where the optimiser likes to go, so each recommendation is re-run through the full ASTRA simulator.

Source: main_pipeline.ipynb (cell 5: predict_surrogate_feasible_batch, optuna objective)

AI-Based CT Acquisition Parameter Prediction | simulation-based study

13 of 18

13

RESULTS

Inverse-recommendation validation (6 tasks)

Optuna reaches the target CNR most closely; the heuristic is both less accurate and slower to scan.

Source: Docs/inverse_recommendation_validation.csv (brute force: 3 of 6 tasks, 4.47% / 8.60 s, omitted)

AI-Based CT Acquisition Parameter Prediction | simulation-based study

14 of 18

14

RESULTS

Pareto frontier case studies

Surrogate frontiers, re-simulated, match brute-force frontiers in hypervolume.

Case

HV surrogate (pred.)

HV surrogate (re-sim)

HV brute force

Max CNR

Frontier CNR err

A: Al, 8 mm cylinder

3,115

3,088

3,108

29.4

4.40%

B: Steel, 15 mm cylinder

686

670

669

6.97

3.62%

Fig. 7 Pareto frontiers: surrogate vs brute force (project output).

Takeaway

Dominance: A beats B if CNR is higher and time not longer. Many non-dominated trade-offs remain.

Source: Docs/pareto_frontier_metrics.csv

AI-Based CT Acquisition Parameter Prediction | simulation-based study

15 of 18

15

RESULTS

Worked recommendation: target CNR 22, budget 35 s

The AI-recommended scan nearly meets the target in a fraction of the maximal-scan time.

Setting

kV

mAs

N

Cu mm

SNR

CNR

Time s

A. Under-exposed trial

90

0.15

75

0.20

8.0

5.6

5.4

B. AI-recommended (Optuna)

213.8

0.385

75

0.23

26.7

21.0

8.9

C. Naive maximal

210

1.5

360

1.0

47.8

34.8

115.4

Fig. 8 Side-by-side reconstructions of settings A, B, C (project output).

Takeaway

B misses the CNR target by 4.6% but uses 8% of C’s scan time.

Source: Docs/worked_demo_comparison.csv; main_pipeline.ipynb cell 6 (target 22, budget 35 s)

AI-Based CT Acquisition Parameter Prediction | simulation-based study

16 of 18

16

SCOPE

Limitations and scope of validity

What the study demonstrates is internal consistency within a simulator, not scanner performance.

  • Simulation only: no physical-scanner data; the spectrum is a Kramers + three-line model, not a measured tube spectrum
  • Idealised phantoms: three materials, one inclusion, single central slice; no scatter, detector response or geometric blur
  • Proxies: scan time, dose and edge FWHM are model quantities, not measurements
  • Extrapolation: XGBoost CNR R² = 0.55 in E2; recommendations near range edges are less reliable
  • Recommendations validated in the same simulator, on only 6 tasks (brute force on 3)
  • Search blend weights (50/35/15) are hand-set; evaluated XGBoost differs from the refit one used in the search
  • Starvation classifier trained on all 3,150 runs, not evaluated on held-out data
  • Result tables label search methods “XGBoost” although a blend is used

AI-Based CT Acquisition Parameter Prediction | simulation-based study

17 of 18

17

CONCLUSIONS

What was demonstrated, and what remains

A simulation-validated recipe for fast, constraint-aware CT parameter recommendation.

Demonstrated (in simulation)

Surrogates predict SNR/CNR with R² up to 0.985; Optuna recommendations average 2.94% CNR error after re-simulation; frontiers match brute force on two cases.

Not yet demonstrated

Real-scanner transfer, other materials, calibrated dose/resolution, robustness to scatter and detector effects.

  • Next: measured spectra and phantom scans for calibration
  • Next: uncertainty-aware (GP) search; fair held-out model selection
  • Next: 3D phantoms and more materials

AI-Based CT Acquisition Parameter Prediction | simulation-based study

18 of 18

18

REFERENCES

References

Project results are attributed to project-generated data; sources below support methods and data.

Berger M.J. et al., XCOM: Photon Cross Sections Database, NIST Standard Reference Database 8 (XGAM).

Feldkamp L.A., Davis L.C., Kress J.W., Practical cone-beam algorithm, J. Opt. Soc. Am. A 1, 612–619 (1984).

van Aarle W. et al., Fast and flexible X-ray tomography using the ASTRA toolbox, Opt. Express 24, 25129–25147 (2016).

McKay M.D., Beckman R.J., Conover W.J., Comparison of three methods for selecting values of input variables, Technometrics 21, 239–245 (1979).

Chen T., Guestrin C., XGBoost: A scalable tree boosting system, Proc. KDD, 785–794 (2016).

Breiman L., Random forests, Machine Learning 45, 5–32 (2001).

Rasmussen C.E., Williams C.K.I., Gaussian Processes for Machine Learning, MIT Press (2006).

Akiba T. et al., Optuna: A next-generation hyperparameter optimization framework, Proc. KDD (2019).

Pedregosa F. et al., Scikit-learn: Machine learning in Python, J. Mach. Learn. Res. 12, 2825–2830 (2011).

AI-Based CT Acquisition Parameter Prediction | simulation-based study