PaperlayerView exact revision on arXiv

Source-linked AI summary

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, Ehsan Abadi

arXiv:2609.13309v1cs.CVcs.AIcs.LG

TL;DR

Exhaustive CT protocol testing is impractical, and limited sampling can risk suboptimal selection. The paper combines virtual imaging trials with reinforcement learning for patient-specific protocol selection; on held-out patients, eight simulations recovered 98.2% of the exhaustive-search oracle objective.

  • Problem

    Exhaustive testing of CT parameter combinations is often impractical, and limited sampling risks converging to suboptimal protocols.

    Source: §I · pp. 1–2

  • Method

    The framework uses virtual imaging trials to train a reinforcement-learning policy conditioned on patient localizer information and selecting protocols from surrogate-scored candidates.

    Source: G. Localizer Encoding via Vision… · p. 3, 2) Baseline Selectors · p. 5, §IV · p. 9

  • Results

    98.2% of the exhaustive-search oracle was recovered with up to eight simulations per patient, while surrogate scoring alone recovered 89.7%.

    Source: §IV · p. 10

  • Takeaways & Limitations

    The policy can select CT protocols from a patient’s localizer with no patient-specific simulation or with up to eight online protocol simulations.

    Source: §V · p. 10

    Evaluation used a limited cohort of human models with liver lesions, leaving broader inter-patient variability for further characterization.

    Source: §IV · p. 10

Abstract

from arXiv · show

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.

I. INTRODUCTION

CT protocol optimization must balance diagnostic image quality and radiation dose, but exhaustive testing is impractical and physical phantoms lack patient-like anatomical and pathological variability. This study combines virtual imaging trials with reinforcement learning and evaluates liver-lesion protocol optimization using computational human models.

I. INTRODUCTION

  • Exhaustive testing across protocol combinations is often infeasible, while physical phantoms lack the anatomical and pathological variability of real patients.Limited sampling of the parameter space can increase the risk of converging to suboptimal protocols.Source: §I · p. 1
  • The framework combines virtual imaging trials with reinforcement learning to optimize CT protocols beyond limited parameter sampling.The study benchmarks the framework for liver-lesion detection.Source: §I · p. 2, Figure 1 · p. 2

A. Study Framework

  • The study framework uses virtual imaging trials to evaluate protocols by balancing lesion detectability against radiation dose and tests performance against an exhaustive-search oracle.The framework is summarized as a study workflow in Fig. 2.Source: A. Study Framework · p. 2, Figure 2 · p. 2

B. Computational Phantoms with Liver Lesions

  • The cohort comprised 63 computational human models with liver lesions, including one to six randomly placed lesions per model.The models represented diverse body habitus and lesion characteristics.Source: B. Computational Phantoms with Liver… · p. 2

C. Image Acquisition

The study evaluates CT acquisition and reconstruction settings using a task-based lesion-detection objective. It estimates detectability from reconstructed-image resolution and noise properties, with a hypothetical lesion task specified in advance.

C. Image Acquisition

  • Acquisitions used a validated CT simulator, and the studied tube-voltage and tube-current settings are summarized in Table I.The simulations were performed at Oak Ridge National Laboratory using the Summit supercomputer.Source: C. Image Acquisition · p. 2, Table I · p. 3

D. Image Reconstruction

  • Sinograms were reconstructed with filtered backprojection, with filter kernel and f50 controlling filter shape and image sharpness.Reconstructed pixel sizes of 0.5 and 1.0 mm corresponded to 1024 and 512 reconstruction matrices, respectively.Source: D. Image Reconstruction · p. 2

E. Task-Based Image Quality

  • Lesion-detection image quality was quantified with d′ using spatial resolution and noise measurements from reconstructed images.The task function represented a hypothetical 4.2 mm spherical lesion with 43.8 HU contrast; MTF came from the skin-to-air edge response and NPS from 100 soft-tissue regions of interest.Source: E. Task-Based Image Quality · pp. 2–3

F. Dose-Aware Objective

The study scores each patient–protocol pair with an objective that trades lesion detectability against a tube-current dose surrogate. It defines the oracle as the protocol maximizing that objective over the full action space.

F. Dose-Aware Objective

  • The dose-aware objective combines normalized patient-specific detectability with a weighted, normalized tube-current penalty.Larger λ favors lower dose, while smaller λ favors higher detectability; λ = 0.18 was used as the balanced operating point for the main comparisons.Source: F. Dose-Aware Objective · p. 3
  • The oracle protocol is the full-action-space protocol with the highest objective, computed by exhaustive search as an evaluation reference.The policy-training reward was the objective of the visited protocol at each step.Source: F. Dose-Aware Objective · p. 3, Figure 3 · p. 3

G. Localizer Encoding via Vision Transformers

The policy uses an AP CT localizer acquired before the diagnostic scan to condition protocol selection on patient-specific information. A pretrained vision transformer supplies image features that are projected to a compact embedding for downstream selectors.

G. Localizer Encoding via Vision Transformers

  • The AP localizer is a pre-scan patient image that captures body size, shape, and anatomy for patient-specific protocol selection.The localizer was generated for each human model with the same validated CT simulator used for tomographic acquisitions.Source: G. Localizer Encoding via Vision… · p. 3
  • A pretrained Vision Transformer served as a fixed feature extractor because the cohort was too small to train an image encoder from scratch without overfitting.The grayscale localizer was replicated across three input channels, and positional embeddings were interpolated to accommodate its elongated image shape.Source: G. Localizer Encoding via Vision… · p. 4
  • A trained projection MLP maps the transformer’s 768-dimensional CLS-token embedding to an L2-normalized 128-dimensional patient embedding.The projection was trained on training patients using an auxiliary objective that included within-patient protocol ranking by d′.Source: G. Localizer Encoding via Vision… · p. 4
  • The downstream selectors use a six-component PCA projection fit on training patients, retaining 93% of training-set embedding variance.The held-out patients were excluded from embedding training, surrogate fitting, policy optimization, hyperparameter selection, and baseline training.Source: G. Localizer Encoding via Vision… · p. 4

H. Data Imputation with a Random-Forest Surrogate

A random-forest surrogate filled missing patient–protocol detectability values, allowing the training objective to be supplied across the protocol space. Its training used simulated cells from training patients, while held-out patients were reserved for oracle evaluation.

H. Data Imputation with a Random-Forest Surrogate

  • 85.9% (20,505) of 23,868 training combinations were simulated; a random-forest surrogate completed the remaining 14.1% (3,363).The combinations span 51 training patients and 468 protocols.Source: H. Data Imputation with a… · p. 4
  • The surrogate predicts d′ from protocol features and a six-dimensional patient-localizer embedding, using ordinal inputs where interpolation is needed and one-hot inputs for unordered axes.The policy instead uses one-hot encoding for every axis because it selects among discrete levels rather than imputing values.Source: H. Data Imputation with a… · p. 4
  • The forest was fit only on simulated training-patient cells, and its predictions were used to select protocols rather than to calculate reported recovery.All 12 held-out test patients retained complete 468-protocol simulations for oracle evaluation and were excluded from surrogate fitting.Source: H. Data Imputation with a… · p. 4

I. Protocol Selectors

Protocol selection is framed as an eight-step budgeted search in which a localizer-conditioned PPO policy proposes protocols. The candidate list can be scored without patient-specific simulation or verified with up to eight simulations, and the selectors are compared with bandit and rule-based baselines.

1) Localizer-Conditioned PPO Selector:

  • The PPO state combines a six-component localizer embedding, the current protocol’s factored one-hot encoding, and the normalized step index; each action selects a complete protocol from 468 combinations.Training rewards use the dose-aware objective, with simulated d′ values where available and random-forest imputation for missing combinations; the search budget is T = 8 steps.Source: 1) Localizer-Conditioned PPO Selector · p. 4
  • 32 policy rollouts generate candidate protocols without observing held-out objectives, and surrogate scores rank them for either direct offline use or online verification of up to eight candidates.Online selection uses the highest true objective among the simulated candidates; offline selection uses the top surrogate-ranked protocol.Source: 1) Localizer-Conditioned PPO Selector · p. 5
  • The contextual bandit ranks all 468 protocols in a one-shot decision, providing a comparison with PPO’s search strategy using the same patient representation and surrogate, with or without the localizer.The exhaustive-search oracle is an evaluation upper bound, not a deployable competitor.Source: 2) Baseline Selectors · p. 5

2) Baseline Selectors:

  • Reference selectors include uniform random choice, a cohort-fixed protocol, a localizer-derived WED dose rule, and WED-kNN based on body size.The WED rule holds reconstruction settings fixed, while WED-kNN selects the neighbor patients’ oracle-optimal protocol by majority vote.Source: 2) Baseline Selectors · p. 5

J. Evaluation and Metrics

Evaluation uses objective recovery against the exhaustive-search oracle, alongside secondary detectability and dose metrics. The exhaustive search shows substantial patient-specific variation in lesion visibility and a dose–detectability trade-off controlled by the penalty weight.

J. Evaluation and Metrics

  • Recovery is the selected protocol’s objective divided by the exhaustive-search oracle objective, aggregated primarily as a ratio of summed objectives across patients.This ratio-of-means weights patients by their oracle objective; an unweighted per-patient mean is also reported, with d′ and selected dose as secondary metrics.Source: J. Evaluation and Metrics · p. 5

A. Exhaustive Search

  • Across three held-out cases, Fig. 4 contrasts each patient’s minimum- and maximum-d′ protocols, with maximum d′ ranging from 3.68 to 6.61.The cases span different body habitus and lesion sizes, and the detectability-optimal protocol is patient-specific.Source: Figure 4 · p. 6, A. Exhaustive Search · p. 6
  • 60 ± 2% lower d′ followed lowering tube current from 150 to 25 mAs, while halving the matrix from 1024 to 512 reduced d′ by 34±14%.The reconstruction kernel also affected d′; slice thickness and f50 had smaller effects, while increasing tube voltage from 100 to 120–140 kV increased d′ by approximately 10±10% and 10±13%.Source: A. Exhaustive Search · p. 6, Figure 5 · p. 6
  • Raising λ reduced mean selected tube current from 127 ± 39 to 95 ± 45 mAs and mean detectability from d′ = 4.05 ± 1.96 to 3.59 ± 2.13.These endpoints span λ = 0.14 to 0.20; λ = 0.18 was selected as the balanced operating point.Source: A. Exhaustive Search · p. 6

B. Performance of Random Forest Surrogate

The random-forest surrogate was assessed against simulated detectability and objective rankings, with calibration checks near the top of its predictions. The supplied dose–detectability comparison identifies the penalty weight used for the main comparisons.

B. Performance of Random Forest Surrogate

  • 0.90 was the Spearman correlation between predicted and true dose-aware objective J rankings across all 468 protocols.For d′, the surrogate achieved Spearman ρ = 0.93–0.94 and top-1 protocol-recovery consistency of 0.90–0.94 on held-out patients.Source: B. Performance of Random Forest… · p. 6
  • As λ rose, mean selected detectability fell from 4.05 to 3.59 and mean selected tube current from 127 to 95 mAs; λ = 0.18 was the balanced operating point.The comparison uses patient-specific oracle selections on the 51-patient cohort.Source: Figure 6 · p. 7
  • −0.005 Bland–Altman bias and −0.002 error within the predicted top decile indicate strong top-tail calibration.The 95% limits of agreement were ±0.10; this ranking fidelity supports filling unsampled training cells and scoring deployment candidates.Source: B. Performance of Random Forest… · p. 7

C. Performance of the Proposed PPO Agent

The proposed PPO agent approached the exhaustive-search oracle with a small simulation budget and showed higher zero-simulation recovery when conditioned on the patient localizer. Results also compare its patient representation and performance with other selectors.

C. Performance of the Proposed PPO Agent

  • The localizer-conditioned PPO policy recovered 10.7 percentage points more than its localizer-blind counterpart, while PCA-6 exceeded the full embedding by 1.8 points.The label-free ViT embedding trailed the d′-supervised PCA-6 representation by 4.9 points even at six dimensions.Source: C. Performance of the Proposed PPO Agent · p. 7
  • The contextual bandit reached 86.4% recovery with the localizer and 82.2% without it, while scoring all 468 protocols by surrogate.PPO committed after scoring only the protocols visited by its rollouts, about 15.3 distinct protocols on average.Source: C. Performance of the Proposed PPO Agent · p. 7
  • Offline PPO exceeded random selection, the cohort-fixed protocol, WED/AEC, and WED-kNN in point estimates, but its margin over the fixed protocol was not significant.From two simulations onward, PPO significantly exceeded the fixed protocol, reaching p = 0.004 at eight simulations.Source: C. Performance of the Proposed PPO Agent · p. 7, Figure 7 · p. 7
  • 98.2% oracle recovery was achieved with up to eight simulations per patient, while offline surrogate scoring alone recovered 89.7%.The cohort mean was 7.5 simulations per patient because some rollouts proposed fewer than eight distinct candidates.Source: C. Performance of the Proposed PPO Agent · p. 8
  • At eight simulations, committed protocols averaged 106.8 mAs, close to the oracle’s 107.1 mAs, rather than defaulting to maximum tube current.Random search reached 64.8 ± 5.0% recovery at the same eight-evaluation budget.Source: C. Performance of the Proposed PPO Agent · p. 8

D. Patient-Level Results and Selected-Protocol Trade-offs

Patient-level results show that eight simulations brought recovery near the oracle across held-out patients, with the largest gains among patients with the weakest offline recovery. Localizer conditioning preserved patient-specific choices.

D. Patient-Level Results and Selected-Protocol Trade-offs

  • Across held-out patients, localizer-conditioned PPO exceeded localizer-blind PPO by 10.7 percentage points and PCA-6 exceeded the full embedding by 1.8 points.The localizer benefit was concentrated in lower-oracle or near-floor patients, where the blind policy more often selected low-detectability protocols.Source: D. Patient-Level Results and… · p. 8
  • Without the localizer, the surrogate’s protocol ranking converged to one global choice: 150 mAs for every patient, an 88% dose increase.The localizer-conditioned scorer instead maintained per-patient rankings.Source: D. Patient-Level Results and… · p. 8
  • With eight simulations, the weakest offline cases improved substantially: pt157 rose from 70.2% to 99.1%, pt200 from 55.8% to 93.2%, and pt163 from 58.4% to 92.9%.The three weakest offline cases were the largest patients by WED; the largest-versus-smallest WED recovery gap narrowed from 21.7 to 3.8 points.Source: D. Patient-Level Results and… · p. 9
  • The committed tube current averaged 106.8 mAs, close to the oracle’s 107.1 mAs.Across eight simulations, tube current or kernel changed for most patients, moving dose toward the oracle.Source: D. Patient-Level Results and… · p. 9
  • Eight simulations narrowed the patient-level recovery range from 55.8%–100.0% offline to 92.9%–100% online.The reported patient-level 95% confidence interval narrowed from 14.9 to 2.6 percentage points.Source: D. Patient-Level Results and… · p. 9

IV. DISCUSSION

The discussion presents a two-mode policy that approximates exhaustive search at low evaluation cost and interprets its behavior through CT physics and dose trade-offs. Its evaluation remains bounded by virtual data, a limited cohort, and a finite protocol space.

IV. DISCUSSION

  • Higher tube current improved d′ by suppressing noise; with fixed lesion contrast, higher voltage modestly improved d′, while softer kernels and lower filter cutoffs improved low-contrast detectability but reduced spatial resolution.The voltage trend may differ for actual lesions, where lower tube potential may increase lesion-to-parenchyma contrast.Source: §IV · p. 9
  • The dose penalty rewards higher detectability only when its gain justifies added tube current; λ = 0.18 was a balanced operating point, not a universal clinical optimum.Different λ values may be selected according to dose constraints and imaging-task priorities.Source: §IV · pp. 9–10
  • 89.7% of the exhaustive-search oracle was recovered without patient-specific simulation, and up to eight candidate simulations raised recovery to 98.2% at under 2% of the evaluation cost.At the same eight-simulation budget, random search reached 64.8 ± 5.0%; the policy's bounded candidates also avoid the lowest-recovery protocols in the full grid.Source: §IV · p. 10
  • The evaluation used a limited cohort with liver-lesion tasks, a tube-current dose surrogate, and a finite action space; prospective evaluation is required before clinical translation.The action space omitted automatic tube-current modulation, pitch, iterative and deep-learning reconstruction, and three reconstruction parameters never varied; prospective work should use clinical localizers, vendor-specific acquisition models, and reader-calibrated task thresholds.Source: §IV · p. 10

V. CONCLUSION

The study integrates virtual imaging trials with reinforcement learning to optimize CT protocols without exhaustive search. A localizer-conditioned policy recovered 89.7% of the oracle without patient-specific simulation and 98.2% when up to eight protocols could be simulated per patient.

V. CONCLUSION

  • 89.7% of the exhaustive-search oracle was recovered with no patient-specific simulation, and 98.2% was recovered when up to eight protocols per patient could be simulated online.The reported values are seed SD 0.5 and 95% CI 80.9–95.7 for the former, and seed SD 1.5 and 95% CI 96.6–99.2 for the latter.Source: §V · p. 10
  • The framework integrates virtual imaging trials with reinforcement learning to optimize CT protocols without exhaustive search.Source: §V · p. 10
  • The results support applying a policy trained on virtual imaging data to a new patient using their localizer image alone.The passage characterizes this application as efficient.Source: §V · p. 10

Related papers

Loading 2609.13309v1…