Abstract
Protocol optimization is critical in Computed Tomography (CT) for achieving desired diagnostic image quality while minimizing radiation dose. Due to the inter-effect of influencing CT parameters, traditional optimization methods rely on the testing of exhaustive combinations of these parameters. This poses a notable limitation due to the impracticality of exhaustive parameter testing. This study introduces a novel methodology leveraging Virtual Imaging Trials (VITs) and reinforcement learning to more efficiently optimize CT protocols. Computational phantoms with liver lesions were imaged using a validated CT simulator and reconstructed with a novel CT reconstruction Toolkit. The optimization parameter space included tube voltage, tube current, reconstruction kernel, slice thickness, and pixel size. The optimization process was done using a Proximal Policy Optimization (PPO) agent which was trained to maximize the Detectability Index (d’) of the liver lesion for each reconstructed image. Results showed that our reinforcement learning approach found the absolute maximum d’ across the test cases while requiring 79.7% fewer steps compared to an exhaustive search, demonstrating both accuracy and computational efficiency, offering a efficient and robust framework for CT protocol optimization. The flexibility of the proposed technique allows for use of varying image quality metrics as the objective metric to maximize for. Our findings highlight the advantages of combining VIT and reinforcement learning for CT protocol management.
Keywords: Computed tomography, Virtual imaging trials, Reinforcement learning, Black-box optimization
INTRODUCTION
Computed Tomography (CT) protocols define a set of instructions, guidelines, and expectations to complete a specific type of radiological examination or procedure. These protocols include the collection of settings that fully describe the CT acquisition and reconstruction parameters [1]. As part of a robust CT protocol management program, CT protocols are routinely reviewed and optimized to achieve desired diagnostic image quality metrics while keeping radiation dose as low as possible.
Protocol optimization requires imaging at various acquisition and condition settings. Therefore, such experiments are impractical with human subjects. Alternatively, several manual approaches have been developed to optimize CT parameters to meet specific image quality or dose requirements using physical phantoms [2, 3]. However, physical phantoms do not represent a diverse patient population with target diseases. Further, it is infeasible to optimize a CT parameter space with an exhaustive search approach, given the substantial number of combined parameters that need to be considered. These limitations can be overcome using realistic virtual imaging trials (VITs) [4], where a diverse patient population can be imaged multiple times with various imaging protocols. VITs have been used extensively in recent years to optimize image quality [5–7] using a uniform sampling approach.
Given that an exhaustive search is computationally infeasible, there is a need to develop a black-box optimization approach that leverages the VIT framework and enables efficient and accurate CT protocol optimization. In this study, we aimed to develop such a framework.
METHODS
Computational Models with Liver Lesions
Sixty-six (thirty-seven male, twenty-nine female) anthropomorphic, computational XCAT phantoms were developed with varied severity of liver lesions [8]. The body habitus for each of the phantoms was based on clinical CT data. The average Body Mass Index (BMI) and age for the phantoms were 26.8 +/− 5.4 kg/m2 and 45.7 +/− 17.7 years, respectively. The liver anatomy for each of the human models included detailed hepatic vasculature developed using a physics-based algorithm [9]. One to six liver lesions of varying sizes (3.9 to 14.9 mm) were placed randomly within the liver.
Imaging Protocols
The human models were imaged using a validated CT simulator (DukeSim) to generate projection images [10] and were reconstructed using the open-source MCR Toolkit [11]. DukeSim was set to model a vendor-neutral scanner with geometric and physical attributes that represent modern CT scanners [8] and the DukeSim acquisitions were performed using the supercomputer Summit at Oak Ridge National Laboratory. The studied parameter space was tube voltage (kV) and current (mAs) on the acquisition side, and kernel shape and sharpness, slice thickness, and matrix size on the reconstruction side. These parameters were discretized with a clinically relevant range, summarized in Table 1.
Table 1.
Acquisition and reconstruction parameters.
| Parameter Type | Parameter Name | Parameter Space |
|---|---|---|
| Acquisition | Tube Voltage | 100, 120, 140 kV |
| Tube Current | 25, 80, 150 mAs | |
| Acquisition Mode | Helical | |
| Reconstruction | Filter Kernel | Ram-lak, Cosine, Smooth, Sharp, Enhancing |
| Filter f50 | 0.4, 0.6, 0.8 lp/mm | |
| Slice Thickness | 0.5, 1.0 mm | |
| Reconstructed Pixel Size | 0.5, 1.0 mm |
Optimization Algorithm Development and Testing
Figure 1 shows our overall optimization framework. We developed a reinforcement learning algorithm consisting of a Proximal Policy Optimization (PPO) agent and an acquisition and reconstruction environment to determine the optimal parameters. The input to the agent was a patient attribute vector which included the patient’s BMI and sex, and the neural network architecture consisted of 2 layers with 64 units per layer. We defined the reward (to be maximized) as the detectability index (d’) of the liver lesion located within the liver parenchyma of each phantom. We trained the PPO agent for 300 steps across 3 patients. For each step, the agent had the ability to acquire and reconstruct images for a single patient with any combination of parameters listed in Table 1. The agent’s goal was to learn actions that maximize the reward, i.e., maximize d’, for each episode. Although d’ was used to define the reward in this study, it can be altered to other quality metrics depending on the optimization task.
Figure 1.
Flowchart of the forward model, consisting of image acquisition, image reconstruction, and image analysis components. Inputs are the XCAT phantom and the acquisition and reconstruction parameters outlined in Table 1 and the outputs are the reconstructed image and detectability index (d’).
Algorithm Analysis
For each patient, we performed a full sweep of the parameter space to obtain the d’ associated with each parameter combination. To evaluate the effectiveness of the reinforcement learning model, the error between the d’ values obtained from the reinforcement learning approach were compared against the maximum d’ obtained from the exhaustive search.
RESULTS
Figure 2 shows a qualitative comparison of images acquired under imaging conditions that yielded the minimum d’ and maximum d’ values for three human models across the studied parameter space (Table 1). Figure 3 shows the learning curve of the proposed reinforcement learning algorithm for three cases in the dataset. For each case, the reinforcement learning algorithm found the absolute maximum d’. The reinforcement learning approach required 100 steps to find the local minimum, whereas the exhaustive search approach required 468 steps; thus, our algorithm required 79.7% fewer steps to determine the parameters required to find the imaging condition that maximizes d’.
Figure 2.
Comparison between minimum d’ and maximum d’ for three liver lesion cases. For each case, an exhaustive search of the parameter space (Table 1) was performed to determine the minimum and maximum d’.
Figure 3.
Learning curves for three cases using the proposed reinforcement learning algorithm. The algorithm predicts the action (i.e., set of acquisition and reconstruction parameters) that provides the absolute maximum d’ for each case while requiring fewer steps than an exhaustive search methodology.
NEW WORK
We developed a novel methodology for determining CT parameters to optimize image quality, leveraging a realistic virtual imaging trial and a reinforcement learning approach. This work has not been submitted for publication or presentation elsewhere.
CONCLUSION
Using a VIT framework consisting of computational human models, a validated CT simulator (DukeSim), and the open-source MCR Toolkit, we developed a methodology for determining CT parameters to optimize a given image quality metric. The use of a VIT framework enabled us to validate our algorithm’s predictions against a ground truth benchmark. This study demonstrated the efficiency and accuracy of our novel reinforcement learning approach and the benefits of the proposed algorithm over a traditional exhaustive search approach.
ACKNOWLEDGEMENT
This work was partially supported by the National Institutes of Health (R01HL155293 & P41EB028744).
REFERENCES
- [1].“AAPM Medical Physics Practice Guideline 1.a: CT Protocol Management and Review Practice Guideline,” Journal of Applied Clinical Medical Physics, 14(5), 3–12 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
- [2].Martini K, Moon JW, Revel MP et al. , “Optimization of acquisition parameters for reduced-dose thoracic CT: A phantom study,” Diagn Interv Imaging, 101(5), 269–279 (2020). [DOI] [PubMed] [Google Scholar]
- [3].Zarb F, Rainford L, and McEntee MF, “Developing optimized CT scan protocols: Phantom measurements of image quality,” Radiography, 17(2), 109–114 (2011). [Google Scholar]
- [4].Abadi E, Segars WP, Tsui BMW, et al. , “Virtual clinical trials in medical imaging: a review,” J. Med. Imag 7(4) 042805 (11 April 2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- [5].Barufaldi B, Maidment ADA, Dustler M et al. , “Virtual Clinical Trials in Medical Imaging System Evaluation and Optimisation,” Radiat Prot Dosimetry, 195(3-4), 363–371 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
- [6].Vancoillie L, Cockmartin L, Lueck F et al. , “Optimized signal of calcifications in wide-angle digital breast tomosynthesis: a virtual imaging trial,” Eur Radiol, (2024). [DOI] [PubMed] [Google Scholar]
- [7].Abadi E, Segars WP, Harrawood B et al. , “Virtual clinical trial for quantifying the effects of beam collimation and pitch on image quality in computed tomography,” J Med Imaging (Bellingham), 7(4), 042806 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
- [8].Abadi E, Segars WP, Felice N et al. , “AAPM Truth-based CT (TrueCT) reconstruction grand challenge,” Med Phys, 2025, 1–13. [DOI] [PMC free article] [PubMed] [Google Scholar]
- [9].Sauer TJ, Bejan A, Segars P et al. , “Development and CT image-domain validation of a computational lung lesion model for use in virtual imaging trials,” Med Phys, 50(7), 4366–4378 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
- [10].Abadi E, Harrawood B, Sharma S et al. , “DukeSim: A Realistic, Rapid, and Scanner-Specific Simulation Framework in Computed Tomography,” IEEE Trans Med Imaging, 38(6), 1457–1465 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
- [11].Clark DP, and Badea CT, “MCR toolkit: A GPU-based toolkit for multi-channel reconstruction of preclinical and clinical x-ray CT data,” Med Phys, 50(8), 4775–4796 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]