Task-Based CT Protocol Optimization Using Reinforcement Learning and Virtual Imaging Trials

Jiaqi Zou; David Fenwick; Vahid Tarokh; Nicholas Felice; Jayasai Rajagopal; Anuj Kapadia; Ehsan Samei; Navid NaderiAlizadeh

cs.CVSep 10, 2026

Task-Based CT Protocol Optimization Using Reinforcement Learning and Virtual Imaging Trials

Authors: Jiaqi Zou, David Fenwick, Vahid Tarokh, Nicholas Felice, Jayasai Rajagopal, Anuj Kapadia, Ehsan Samei, Navid NaderiAlizadeh, +1 more

Organizations: Department of Electrical and Computer Engineering, Duke University, Durham, NC 27705 USA · Department of Radiology and Medical Physics Graduate Program, Duke University, Durham, NC 27705 USA · Oak Ridge National Laboratory, Oak Ridge, TN 37830 USA · Departments of Electrical and Computer Engineering, Radiology, Biomedical Engineering, Physics, and Medical Physics Graduate Program, Duke University, Durham, NC 27705 USA · Department of Biostatistics and Bioinformatics, Duke University, Durham, NC 27705 USA · Departments of Electrical and Computer Engineering, Radiology, and Medical Physics Graduate Program, Duke University, Durham, NC 27705 USA

Abstract

Protocol optimization in computed tomography (CT) aims to improve diagnostic image quality while reducing radiation dose, but the interdependence of acquisition and reconstruction parameters makes exhaustive testing impractical. We propose a virtual imaging trial framework with reinforcement learning for efficient CT protocol optimization. Sixty-three computational human models with liver lesions were imaged using a validated CT simulator across 468 combinations of acquisition and reconstruction parameters, including tube voltage, tube current, reconstruction kernel, slice thickness, and pixel size. The optimization objective balanced liver lesion detectability, quantified by detectability index d-prime, against radiation dose. A Proximal Policy Optimization agent was trained and conditioned on patient-specific CT localizer embeddings derived from a pretrained vision transformer. On held-out patients, evaluating only 8 protocols per patient, about 2% of exhaustive testing, recovered 98.2% of the exhaustive-search oracle objective. With no patient-specific simulation, surrogate scoring alone achieved 89.7% recovery. Conditioning on the localizer improved zero-simulation recovery by 10.7 percentage points over the localizer-blind policy (paired 95% CI 2.9-19.5; p=0.02). These results show that the proposed framework can substantially reduce exhaustive protocol testing while enabling task-based, dose-aware protocol selection before the diagnostic scan.

Figures & tables

Fig. 1: CT protocol optimization framework using virtual imaging trials (VITs) as the forward model.

Fig. 2: Overview of the proposed framework. The forward model generates CT images of human models for any protocol acquisition-reconstruction setting, denoted by

a

. It also generates the patient’s CT localizer image. The objective (

J_{p}(a)

) balances image quality (detectability index

d^{\prime}_{p}(a)

) against radiation dose (tube current

D(a)

), with a fixed weight

\lambda

controlling the trade-off. The subscript

p

denotes the patient. ViT is vision transformer and MLP is multi-layer perceptron. The PPO policy, conditioned on the localizer embedding

\tilde{z}_{p}

and rewarded by

J_{p}(a)

, returns its selected protocol

\hat{a}_{p}

to the simulator, closing the optimization loop.

Parameter Type	Parameter Name	Parameter Space
Acquisition	Tube Voltage	100, 120, 140 kV
	Tube Current	25, 80, 150 mAs
Reconstruction	Filter Kernel	Ram-lak, Cosine, Smooth, Sharp, Enhancing
	Filter $f_{50}$	0.4, 0.6, 0.8 mm -1
	Slice Thickness	0.5, 1.0 mm
	Pixel Size	0.5, 1.0 mm

TABLE I: Acquisition and reconstruction parameters

Fig. 3: Architectures and training of the localizer-conditioned selector (colors as in Fig. 1 ; gray: frozen; green: trainable). The ViT-B/16 backbone is frozen; only the projection MLP is trained, with a within-patient protocol-ranking loss on the training patients so that the embedding orders protocols by

d^{\prime}

within each patient, and the PCA projection is fit on the training patients only. The policy and value networks are trained with the clipped PPO objective (

\gamma=0.99

) under the eight-step episode budget, with the dose-aware objective

J_{p}

as the per-step reward. At deployment, each visited protocol is scored by the random-forest surrogate and the best-seen protocol

\hat{a}_{p}

is committed; the off-line mode does not have access to held-out ground truth.

Fig. 4: Qualitative case studies of protocol impact for three held-out patients spanning body habitus and lesion size. For each patient, the reconstructed axial slice with zoomed lesion inset is shown for the worst protocol in the space (left, minimum $d^{\prime}$ ) and the best protocol (right, maximum $d^{\prime}$ ). Protocol choice moves the worst-case lesion from nearly invisible (minimum $d^{\prime}\leq 0.74$ ) to clearly detectable (maximum $d^{\prime}=3.68$ – $6.61$ ), and the detectability-optimal protocol is patient-specific, motivating localizer-conditioned personalized selection.

Fig. 5: Sensitivity of detectability

d^{\prime}

to (a) acquisition parameters and (b) reconstruction parameters across the 63-patient cohort. Each box shows the per-patient percent change in

d^{\prime}

when one parameter is varied from a common baseline (100 kV, 150 mAs, cosine kernel,

f_{50}=0.4

mm -1 , 1.0 mm slices, 1024 matrix);

n=49

–

59

patients per comparison with both cells simulated. Box: interquartile range across patients; line: median; whiskers:

1.5\times

IQR; outliers omitted.

Fig. 6: Dose–detectability trade-off across the dose-penalty weight

\lambda

, computed from the patient-specific oracle selections of the objective on the 51-patient cohort. Raising

\lambda

monotonically reduces both the mean selected detectability (left axis,

4.05\rightarrow 3.59

) and the mean selected tube current (right axis,

127\rightarrow 95

mAs); error bars:

\pm

1 SD across patients. We set

\lambda=0.18

as the balanced operating point for the main comparisons.

Fig. 7: Off-line (zero-simulation) oracle objective recovery of the deployable selectors on held-out patients at

\lambda=0.18

; this is the operating point reached with no patient-specific simulation. Filled markers: trained with the localizer; open markers: trained without it; the connector length is the localizer gain (

+10.7

pp for PPO,

+4.2

pp for the bandit). Whiskers:

\pm

1 SD (five seeds for PPO, five cross-validation folds for the bandit); the rule-based selectors are deterministic. The patient-level 95% bootstrap CI for the localizer-conditioned PPO point is 80.9–95.7%. The exhaustive-search oracle (

100\%

) is shown dashed and the random floor (

31.0\%

) lies off scale.

XCAT ID	$d^{\prime}$	mAs	$d^{\prime}$	mAs	Rec. (%)	$d^{\prime}$	mAs	Rec. (%)	VIT budgets
	Oracle		Offline (no simulation)			Online (8 simulations)
pt152	7.93	150	7.92 $\pm$ 0.01	150	99.9 $\pm$ 0.2	7.92 $\pm$ 0.01	150	99.9 $\pm$ 0.2	5.6
pt300	6.62	150	4.95 $\pm$ 0.03	80	87.5 $\pm$ 0.6	6.62 $\pm$ 0.00	150	100.0 $\pm$ 0.1	8.0
pt269	5.13	150	4.86 $\pm$ 0.01	150	96.6 $\pm$ 4.1	5.03 $\pm$ 0.13	150	96.7 $\pm$ 4.0	6.8
pt157	4.58	150	2.70 $\pm$ 0.01	80	70.2 $\pm$ 0.2	4.34 $\pm$ 0.48	136 $\pm$ 28	99.1 $\pm$ 1.8	7.2
pt141	4.46	150	3.17	80	92.0	3.68 $\pm$ 0.61	108 $\pm$ 34	95.1 $\pm$ 3.1	8.0

TABLE II: Per-patient results on the 12 held-out test patients (

\lambda=0.18

Simulations	PPO (ours)	Bandit	Random
per patient	(%)	(%)	(%)
0 (off-line)	89.7 $\pm$ 0.5	86.4 $\pm$ 2.1	31.0
2	90.5 $\pm$ 0.7	87.5 $\pm$ 1.5	43.9
4	94.0 $\pm$ 0.4	92.0 $\pm$ 1.5	55.1
6	97.0 $\pm$ 0.8	94.9 $\pm$ 1.3	60.9
8	98.2 $\pm$ 1.5	95.8 $\pm$ 1.3	64.8

TABLE III: Oracle recovery versus on-line VIT budget (

\lambda=0.18

XCAT ID	kV	mAs	kernel	kV	mAs	kernel	kV	mAs	kernel	Rec. off $\rightarrow$ on (%)
	Oracle			Offline (no simulation)			Online (8 simulations)
pt152	120	150	cos	120	150	cos	120	150	cos	99.9 $\rightarrow$ 99.9
pt300	120	150	cos	140	80	smo	120	150	cos	87.5 $\rightarrow$ 100.0
pt269	140	150	smo	120	150	smo	140	150	smo	96.6 $\rightarrow$ 96.7
pt157	120	150	smo	140	80	smo	120	150	smo	70.2 $\rightarrow$ 99.1
pt141	140	150	cos	140	80	smo	140	80	cos	92.0 $\rightarrow$ 95.1

TABLE IV: Per-patient selected protocols (tube voltage / tube current / kernel) and recovery, offline versus online, at

\lambda=0.18

Explore similar work

SAVER: Stochastic Adaptive Variance-Driven Exploration and Reconstruction for Low-Dose Computed Tomography
Jul 4, 2026Shunta Nonaga, Koji Tabata, Junya Honda +3Sparse-View CT Reconstruction Adaptive Sampling
Graduate School of Chemical Sciences and Engineering, Hokkaido University, Kita 13 Nishi 8, Kita-ku, Sapporo, Hokkaido 060-8628, Japan. · Research Center of Mathematics for Social Creativity, Research Institute for Electronic Science, Hokkaido University, Kita 20 Nishi 10, Kita-ku, Sapporo, Hokkaido 001-0020, Japan · Institute for Chemical Reaction Design and Discovery (WPI-ICReDD), Hokkaido University, Kita 21 Nishi 10, Kita-ku, Sapporo, Hokkaido 001-0021, Japan. +7
Agentic Autoresearch for CT Reconstruction
Jul 24, 2026Andreas Maier, Lucas Kachelriess, Siming Bayer +4AI for Science Sparse-View CT Reconstruction
Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Erlangen, Germany · Department Artificial Intelligence in Biomedical Engineering (AIBE), Friedrich-Alexander-Universität Erlangen-Nürnberg (FAU), Erlangen, Germany · Universitätsklinikum Erlangen (UKER), Erlangen, Germany +2
RA-CMF: Region-Adaptive Conditional MeanFlow for CT Image Reconstruction
Apr 28, 2026Md Shifatul Ahsan Apurba, Md Selim, Jin ChenImage Restoration Computed Tomography
Biomedical Informatics and Data Science University of Alabama at Birmingham Birmingham, AL, United States · Computer Science Florida Polytechnic University Lakeland, FL, United States

Jul 4, 2026J/K move · Enter open · S save

SAVER: Stochastic Adaptive Variance-Driven Exploration and Reconstruction for Low-Dose Computed Tomography

Shunta Nonaga, Koji Tabata, Junya Honda, Hiroyuki Kudo+2

Graduate School of Chemical Sciences and Engineering, Hokkaido University, Kita 13 Nishi 8, Kita-ku, Sapporo, Hokkaido 060-8628, Japan. · Research Center of Mathematics for Social Creativity, Research Institute for Electronic Science, Hokkaido University, Kita 20 Nishi 10, Kita-ku, Sapporo, Hokkaido 001-0020, Japan · Institute for Chemical Reaction Design and Discovery (WPI-ICReDD), Hokkaido University, Kita 21 Nishi 10, Kita-ku, Sapporo, Hokkaido 001-0021, Japan.+7

Sparse-View CT Reconstruction Adaptive Sampling Computed Tomography

Open paper AskPDF arXiv

Computed Tomography (CT) is indispensable in clinical diagnostics, yet minimizing radiation dose without compromising image quality remains a critical challenge. Conventional low-dose protocols often rely on fixed, uniform angular sampling, independent of the underlying structural complexity of organs of individual patients. We propose ``Stochastic Adaptive Variance-Driven Exploration and Reconstruction'' (SAVER), an adaptive data acquisition framework that selects projection angles in real-time based on the statistical variance of acquired data. Utilizing a Softmax-based stochastic scheduling scheme with simulated annealing, SAVER prioritizes directions with high structural information while maintaining necessary exploration. Numerical experiments across 8 diverse phantoms demonstrate that SAVER achieves consistently higher reconstruction fidelity than conventional random sampling, particularly for objects with high structural anisotropy. Furthermore, the proposed method exhibits robust performance under significant measurement noise. By dynamically reallocating radiation dose to the most informative projections, SAVER provides a mathematically-grounded approach to maximize diagnostic quality per unit of radiation dose, marking a shift toward sample-dependent, data-driven CT acquisition.

Abstract

Explore similar work

SAVER: Stochastic Adaptive Variance-Driven Exploration and Reconstruction for Low-Dose Computed Tomography

Agentic Autoresearch for CT Reconstruction

RA-CMF: Region-Adaptive Conditional MeanFlow for CT Image Reconstruction