Innovations in artificial intelligence for pet/mr imaging: Application and performance analysis

Abstract

Background

The primary challenges in PET/MR imaging include prolonged scan durations for both PET and MR components and radiation exposure associated with the PET modality. Artificial intelligence (AI)-based techniques offer a promising approach to overcome these limitations.

Objective

This study evaluates the AI-based image enhancement methods integrated into the United Imaging PET/MR system, focusing on improvements in image quality, reduced injection dose, and shortened acquisition duration.

Method

Sixty-three patients underwent ¹⁸F-FDG PET/MR scans using uPMR790 (0.09 ± 0.01 mCi/kg, 5 min/bed, n = 29) and uPMR890 (0.05 ± 0.01 mCi/kg, 2.5 min/bed for PET and accelerated MR protocols, n = 34) with advanced AI-enhanced method. Shortened MR protocols included T1 W and T2 W sequences. Image quality was evaluated subjectively by two physicians and objectively using SNR and artifact ratios.

Results

The AI-enhanced system achieved high-quality PET and MR images despite reduced PET doses and scan durations for both PET and MR components. AI-based reconstruction images showed higher SNR, fewer artifacts, and reduced noise compared to the conventional system.

Conclusions

AI-enhanced PET/MR significantly improves imaging efficiency by reducing PET/MR acquisition durations, lowering radiation dose, and enhancing overall image quality, making it a valuable tool for clinical hybrid imaging.

Keywords

PET/MR imaging low-dose artificial intelligence short-acquisition

Introduction

Integrated Positron Emission Tomography/Magnetic Resonance (PET/MR) technology combines PET molecular imaging with MR multiparametric capabilities, offering a new perspective for molecular functional imaging applications.¹ With the growing clinical use of PET/MR, institutions are enhancing quality control and management while exploring technical advancements to improve image quality and acquisition efficiency.

PET/MR is ideal for early tumor screening in high-risk populations (e.g., BRCA1 mutation carriers) and for improving staging in areas like the brain, liver, and pelvis.^2–5 It is also suited for radiation-sensitive individuals. However, its clinical use faces several challenges: First, the extended imaging duration—averaging 61 ± 18 min—can cause heat buildup, patient discomfort, and bladder filling, which may hinder imaging quality.⁶ Second, balancing MR multiparametric capabilities with shorter acquisition duration is difficult. Julian Kirchner et al. observed that, in whole-body PET/MR for lymphoma, additional T1-weighted contrast-enhanced sequences and DWI sequences are crucial, significantly improving diagnostic accuracy (87% vs. 95%).⁷ Third, MR is more prone to motion artifacts due to extended imaging durations, which may obscure lymph node localization.⁸ Fourth, respiratory-gated MR sequences may be misaligned with PET, eliminating the synchronization advantage of integrated PET/MR. Fifth, MR-based attenuation correction sequences, such as those using breath-hold acquisition, can cause PET Standardized Uptake Value (SUV) quantification inaccuracies in liver lesions, while respiratory-gated methods reduce acquisition efficiency.⁹

In recent years, deep learning neural networks have revolutionized MR and PET imaging acquisition, reconstruction, image synthesis, and artifact detection. Xu et al. implemented AI-assisted compressed sensing techniques (ACS) for MR acceleration, reducing neck tumor imaging duration by 40% without compromising diagnostic efficacy.¹⁰ Küstner et al. developed an automated MR motion artifact recognition method, achieving a 97% accuracy rate in head and neck artifact detection.¹¹ In PET imaging, Chen et al. applied Hyper DPR (a vendor-provided deep progressive reconstruction method), demonstrating improved spatial resolution, Signal-to-Noise Ratio (SNR), and sensitivity in phantom studies.¹² Wang et al. clinically validated this algorithm, showing that it enabled a 60% reduction in radiopharmaceutical dose.¹³ Multiparametric PET/MR imaging with diverse tracers has heightened demands for lesion detection and analysis. Zhang et al. proposed an automated whole-body lesion segmentation method using PET images, enhancing multimodal tumor image analysis.¹⁴ Ahangari et al. utilized synthetic Computed Tomography (sCT) images for PET attenuation correction, improving quantitative accuracy in PET.¹⁵ Moreover, standard surface coils for MR imaging in PET/MR cannot undergo attenuation correction, causing photon scatter and degrading PET image quality. Deller et al. introduced flexible, lightweight coils for PET/MR, reducing quantitative bias by 50%.¹⁶ These technological advancements offer solutions to inherent PET/MR challenges.

This study leverages six advanced technologies —MR ACS accelerated acquisition, PET Hyper DPR, MR DeepRecon reconstruction, sCT-based attenuation correction, μ-map generation, and the low-attenuation flexible coil —to design a low-dose, rapid whole-body PET/MR protocol. This protocol aims to compare the image quality between PET and MR and examine the comprehensive benefits of deep learning neural networks in improving imaging efficiency and diagnostic accuracy across a variety of innovative applications.

Materials and methods

Patient cohorts

This retrospective study included 29 conventional whole-body PET/MR scans collected from Tongji University Affiliated East Hospital from May 2020 to June 2020 (uPMR790, United Imaging Healthcare, Shanghai). Additionally, 34 AI-assisted new generation PET/MR scans were collected from Ruijin Hospital affiliated with Shanghai Jiao Tong University School of Medicine from June 2024 to July 2024(uPMR890, United Imaging Healthcare, Shanghai). All procedures involving human participants were conducted in accordance with the ethical guidelines of the 1964 Helsinki Declaration and national regulations. The study was approved by the Research Ethics Committee of Ruijin Hospital and Shanghai East Hospital.

Deep neural network setup

The PET/MR image acquisition and analysis solution based on deep learning neural networks integrates five neural network-based technologies for image acquisition, reconstruction, and analysis, along with a low-attenuation flexible coil. While motion monitoring and lesion detection technologies do not directly impact image quality, they enhance image acquisition efficiency and automatic analysis capabilities. The low-attenuation flexible coil improves PET quantification accuracy and increases the SNR of MR images. This study quantitatively evaluates the combined performance of six key technologies: MR ACS accelerated acquisition, PET MR DeepRecon reconstruction, sCT-based attenuation correction, μ-map generation, and the low-attenuation flexible coil.

MR ACS accelerated acquisition: Figure 1 illustrates the ACS PET attenuation correction process based on MR images. The ACS network can adopt a conditional generative adversarial network (cGAN) structure, where the generator may use a U-Net or a modified version of it. Four loss functions enhance image quality: pixel-wise loss for visual clarity, frequency domain loss for frequency alignment, perceptual loss for deep feature matching, and adversarial loss for balancing data generation.¹⁷

PET Hyper DPR: This method introduces a deep progressive learning (DPL) strategy for PET image reconstruction, organizing the workflow as a chain of convolutional neural networks (CNNs), each targeting specific image attributes. Feedback networks, trained on high-quality PET images from uEXPLORER, are integrated into the iterative reconstruction process to sequentially control image noise and contrast. Experimental results demonstrate its feasibility in reducing noise and enhancing contrast for improved PET image quality. For further methodological details, please refer to.¹⁸

MR DeepRecon reconstruction: The key to noise reduction in MR imaging using MR DRecon lies in applying a convolutional neural network for residual learning. The network outputs the noise at each pixel rather than the noise-free image itself. It consists of a convolutional input layer with ReLU activation, followed by eight convolutional layers with batch normalization and ReLU, and a convolutional output layer, all without max-pooling to preserve the image dimensions.

sCT-based attenuation correction: The sCT image synthesis network generally adopts a 3D U-Net architecture, taking Dixon in-phase (IP) and out-of-phase (OP) MR images as inputs and producing sCT images as outputs. The network starts with 32 kernels in its first convolutional layer, and both encoder and decoder paths use 3 × 3 × 3 kernels with batch normalization and ReLU activation for improved convergence. After each convolution operation, downsampling is performed with a stride-2 convolutional layer. A critical preprocessing step in this MR-to-sCT conversion is the registration between MR and CT images. Since the MR scanner uses a flat table surface, whereas CT scanners have a curved surface, achieving accurate non-rigid registration between these modalities, especially in the thoracic and abdominal areas, is challenging. In this study, we utilized radiotherapy simulation CT images with a flatbed design, which align well with the MR images. Using mutual information as a registration criterion, we could match anatomical structures between the modalities by continuously adjusting transformation parameters to maximize mutual information, ensuring precise registration.

μ-map generation: For generating the μ-map, most of the μ values are derived by registering MR images to CT images. The μ values obtained using this method include: air 0 cm⁻¹, lung 0.024 cm⁻¹, fat 0.086 cm⁻¹, soft tissue 0.101 cm⁻¹, and bone 0.146 cm⁻¹. Since the grayscale values for the lungs and air around the body are much lower than for soft tissues, threshold-based segmentation is typically used for lung and body contouring on Dixon IP images. However, using a U-Net or an improved U-Net can enhance the segmentation of the lungs and air, overcoming issues caused by low MR signal from gastrointestinal gases and partially calcified tissues.

Low-attenuation flexible coil: The low-attenuation SuperFlex Coil is one of the key technologies, featuring a critical component in the form of a low-noise preamplifier essential for signal reception in MR systems. The RF signals received by local coil elements pass through a preamplifier to enhance SNR, a crucial factor for image quality. Miniaturizing the components within the coil reduces overall weight and size, improving user experience. The SuperFlex Coil, possessing a weight of 2.63 kg, is equipped with a high-performance preamplifier and features a compact circuit design. Notably, a single circuit board thereof weighs merely 6 g, which merely constitutes 21% of that of conventional coil units.

Figure 1.

An illustration of enhanced attenuation correction in hybrid PET/MR imaging using a deep-learning-based continuous μ-map generation framework.

Imaging protocols

The study included 63 participants who underwent whole-body PET/MR scans. Conventional PET/MR involved 4 bed positions, each with a 5-min PET scan and image reconstruction using OSEM. MR sequences included axial single-shot fast spin echo T2-weighted (T2w SSFSE) and T1-weighted water-fat imaging (T1w WFI). AI-assisted PET/MR also used 4 bed positions, but with a reduced PET scan duration of 2.5 min per position and MR DeepRecon reconstruction. MR sequences used ACS-accelerated T2-weighted fast spin echo (T2w FSE) and T1w WFI for faster, high-quality imaging. Table 1 presents the MR sequence parameters in whole-body PET/MR Imaging.

Table 1.

Mr sequence parameters in whole body PET/MR imaging.

Sequence	TR (ms)	TE (ms)	Acceleration factor	Field of view	Matrix size	Slice thickness (mm)	Scan duration (second)
T1 Quick3d WFI	4.86	2.24/3.21	8.58 (ACS)	350*500	232*368	4	12.8
T2 FSE TRA	3950	83.72	3.12 (ACS)	300*450	154*288	4	16
T2 FSE COR FS	6842	117.76	3.01 (ACS)	380*380	192*256	5	14
EPI DWI (b value 800)	5914	67.40	2	320*400	204*256	8	90
T2 FSE SAG FS	5897	73.60	3.06 (ACS)	250*380	126*256	4	12.2
Conventional protocols
T1 Quick3d WFI	4.60	2.40/3.20	6 (uCS)	350*500	206*144	4	21
T2 SSFSE COR	1300	94.20	2	350*450	393*408	6	46
T2 FSE TRA FS	3968	90	2	300*380	337*504	6	156

Note: The upper section presents the MR sequence parameters for an AI-assisted PET/MR system, highlighting advanced acquisition settings for efficient whole-body imaging. The lower section shows the parameters for traditional PET/MR acquisition methods, illustrating differences in scan duration and resolution across techniques.

Qualitative and quantitative evaluation

For the two groups of PET/MR whole-body imaging, both qualitative and quantitative assessments were conducted. Qualitative evaluation was performed by two nuclear medicine physicians who rated image quality based on visual SNR, artifact levels, and uniformity, with scores ranging from 1 (poor) to 5 (excellent). Considering differences in PET attenuation correction methods, PET images were rated separately for the head and body. The μ-map was evaluated for skull, spine, pelvis, ribs, and soft tissues. T1w and T2w images were scored across five regions (head, neck, chest, abdomen, and pelvis), and artifact proportions were calculated as the ratio of the slice number with significant artifacts and overall slices. Quantitative evaluation focused on the ratio of the SUV and standard deviation in the liver for noise assessment and compared the SNR in the liver, brain, and spleen, using the following formula:

S N R_{L i v e r} = \frac{S U V_{m e a n, Liver}}{N o i s e l e v e l}; S N R_{B r a i n} = \frac{S U V_{m e a n, brain}}{N o i s e l e v e l}; S N R_{S p l e e n} = \frac{S U V_{m e a n, spleen}}{N o i s e l e v e l}

Statistical analysis

Statistical analysis was conducted using SPSS software (IBM Corp. Released 2021. IBM SPSS Statistics for Windows, Version 28.0. Armonk, NY: IBM Corp), with a significance threshold of p < 0.05. The demographic information including age and gender were compared using the unpaired t-test. For qualitative comparisons, a Kolmogorov-Smirnov test was first used to compares the cumulative distribution of the data sets to assess normality in the physician ratings. Based on normality, unpaired t-tests (for normally distributed data) or nonparametric Mann-Whitney tests (for non-normal data) were applied to compare the means of the physician ratings and the artifact proportions. For quantitative analysis, the Kolmogorov-Smirnov test was used to evaluate parameter normality. Subsequently, unpaired t-tests or Mann-Whitney tests were used to compare noise levels and SNR in the liver, brain, and spleen between the two groups.

Results

Table 2 presents demographic and imaging data for the study participants, including 29 patients who underwent conventional scanning and 34 patients with AI-assisted scanning, with details of age, gender, PET injection dose, and dose per unit body weight. No significant differences were observed between the groups in terms of age and gender (p > 0.05). However, AI-assisted imaging showed significantly lower PET injection doses and per unit weight doses compared to conventional imaging (p < 0.0001). Figure 2 shows whole-body rapid low-dose sequence display across various sequences from new generation PET/MR imaging. Figure 3 compares PET/MR images obtained with and without MR ACS accelerated acquisition protocol. Even with a shorter MR scan duration, the use of MR ACS accelerated acquisition significantly improves the detection of small lung lesions, demonstrating enhanced sensitivity in identifying these micro-lesions compared to the conventional imaging technique. In Figure 4, the MR DeepRecon reconstruction also enhances the detection of small lesions, with improved contrast and reduced noise, outperforming the OSEM algorithm. The integration of ACS + DeepRecon accelerates MR scanning and delivers images with higher contrast, effectively improving PET image quality and lesion detectability. Visualization of sCT images generated from MR images is shown in Figure 5. The sCT images display higher contrast in bone structures like the ribs, providing more precise attenuation correction during the generation of μ-map. This enhanced accuracy is particularly beneficial for improving the overall quality of PET images.

Figure 2.

Representative PET, mr, and fused PET/mr images from the whole-body rapid low-dose PET/MR imaging. MIP = Maximum Intensity Projection, MPR = Multi-Planar Reformatting, COR = Coronal, FS = Fat Saturation.

Figure 3.

Comparative MR images obtained with and without MR ACS accelerated acquisition technology. The ACS-enhanced approach demonstrates improved sensitivity in detecting small lung lesions (arrow) while significantly reducing MR scan time. The final image represents PET/MR fusion, illustrating the integration of functional and structural information for enhanced diagnostic accuracy.

Figure 4.

Comparison of PET and MR images reconstructed using conventional and AI-enhanced methods. PET images were reconstructed using the conventional OSEM method and the AI-enhanced method, while MR images were reconstructed using the conventional parallel imaging method and the AI-assisted Compressed Sensing (ACS) method. ACS = AI-assisted Compressed Sensing.

Figure 5.

Visualization of sCT images generated from MR images. The sCT images display higher contrast in bone structures like the ribs, providing more precise attenuation correction during the generation of μ-maps.

Table 2.

Information on patients’ demographic.

	Conventional protocols	AI-enhanced	p value
Number of Patients	29	34	/
Gender(female/male)	9/20	12/22	0.7925
Age(year)	57.83 ± 12.19	63.00 ± 10.12	0.0706
Total Dose (mCi)	6.13 ± 1.41	3.14 ± 0.82	<0.0001****
Unit dose per body weight (mCi//kg)	0.09 ± 0.01	0.05 ± 0.01	<0.0001****

Note: The p-value was calculated using the unpaired t-test. **** p < 0.0001.

Qualitative evaluation results

The results of the qualitative imaging assessment by two board-certified nuclear medicine physicians are summarized in Table 3. For PET imaging, AI-enhanced imaging significantly outperformed conventional imaging in both head and body image quality. Similarly, μ-map images displayed better performance in the skull, spine, pelvis, ribs, and soft tissues. In MR, AI-assisted imaging showed advantages in abdominal T1w imaging and in T2w imaging of the neck, chest, abdomen, and pelvis. With the help of ACS rapid imaging and DeepRecon noise reduction, the number of artifact layers in AI-assisted T1w and T2w imaging was significantly lower (as shown in Figure 6).

Figure 6.

Comparison of artifact percentages in T1w and T2w images, as determined by two physicians, between traditional imaging and AI-assisted imaging.

Table 3.

Qualitative evaluation results from 2 physicians.

	Physicians 1			Physicians 2
	Conventional protocols	AI-enhanced	p value	Conventional protocols	AI-enhanced	p value
PET head	4（3 ∼ 5）	5（4 ∼ 5）	0.0014**	4（3 ∼ 5）	5（4 ∼ 5）	0.0066**
PET body	4（3 ∼ 5）	4（3 ∼ 5）	0.0023**	4（3 ∼ 5）	4（3 ∼ 5）	0.0023**
μ-map skull	5（3 ∼ 5）	5（5 ∼ 5）	0.0189*	5（3 ∼ 5）	5（5 ∼ 5）	0.0079**
μ-map spine	3（3 ∼ 3）	5（5 ∼ 5）	<0.0001****	3（3 ∼ 3）	5（5 ∼ 5）	<0.0001****
μ-map pelvis	3（3 ∼ 3）	5（5 ∼ 5）	<0.0001****	3（3 ∼ 3）	5（5 ∼ 5）	<0.0001****
μ-map ribs	3（3 ∼ 3）	5（5 ∼ 5）	<0.0001****	3（3 ∼ 3）	5（5 ∼ 5）	<0.0001****
μ-map soft- tissues	5（3 ∼ 5）	5（4 ∼ 5）	0.0297*	5（3 ∼ 5）	5（4 ∼ 5）	0.0297*
T1w head	5（4 ∼ 5）	5（5 ∼ 5）	0.0991	5（4 ∼ 5）	5（5 ∼ 5）	0.0438*
T1w neck	5（4 ∼ 5）	5（4 ∼ 5）	0.3346	5（4 ∼ 5）	5（4 ∼ 5）	0.3346
T1w chest	5（4 ∼ 5）	5（5 ∼ 5）	0.4727	5（4 ∼ 5）	5（5 ∼ 5）	0.4727
T1w abdomen	4（4 ∼ 5）	5（4 ∼ 5）	<0.0001****	4（4 ∼ 5）	5（4 ∼ 5）	<0.0001****
T1w pelvis	4（4 ∼ 5）	4（4 ∼ 5）	0.2808	4（4 ∼ 5）	4（4 ∼ 5）	0.7389
T2w head	5（4 ∼ 5）	4（4 ∼ 5）	0.0585	4.5（4 ∼ 5）	4（4 ∼ 5）	0.1037
T2w neck	4（3 ∼ 5）	5（4 ∼ 5）	<0.0001****	4（3 ∼ 5）	5（4 ∼ 5）	<0.0001****
T2w chest	5（4 ∼ 5）	5（5 ∼ 5）	0.0013**	5（4 ∼ 5）	5（5 ∼ 5）	0.0013**
T2w abdomen	4（3 ∼ 4）	5（4 ∼ 5）	<0.0001****	4（3 ∼ 4）	5（4 ∼ 5）	<0.0001****
T2w pelvis	4（3 ∼ 5）	4（4 ∼ 5）	0.0070**	4（3 ∼ 5）	4（4 ∼ 5）	0.0041**

Note: The p-value was calculated using the unpaired t-test. **** p < 0.0001.

Quantitative evaluation results

Despite significantly lower PET injection doses and per unit body weight doses in AI-assisted imaging compared to conventional imaging, no statistically significant differences were found in noise levels (Figure 7A) or the signal-to-noise ratios (SNR) for the liver (Figure 7B), brain (Figure 7C), and spleen (Figure 7D).

Figure 7.

Comparison of noise levels and SNR between traditional and AI-enhanced PET imaging.

Discussion

This study demonstrates that AI-assisted imaging significantly improves the quality of PET and MR images in most areas (PET head p = 0.0066, PET body p = 0.0023, MR T2 W neck p < 0.0001, MR T2 W chest p = 0.0013, MR T2 W abdomen p < 0.0001, MR T2 W pelvis p = 0.0070). Quantitative PET analysis shows no significant loss in SNR despite reduced dose. MR T2 W imaging notably reduced artifacts in the body, while head T2 W images occasionally had streak artifacts. MR T1 W imaging benefited from shortened acquisition durations, improving abdominal image quality (Figure 8). AI-assisted sCT attenuation and DeepRecon reconstruction technologies enhanced PET/MR image quality, making short-acquisition low-dose PET/MR comparable to standard protocol.

Figure 8.

Mr T1 W acquisition time was reduced using ACS (from 32 s to 17 s), allowing abdominal imaging with breath-holding, which significantly improved image quality.

In this study, we observed and summarized the innovative application scenarios of various technologies in PET/MR imaging. Specifically, the SuperFlex Coil demonstrated reduced attenuation, with significant advantages in pediatric, breast, prostate, and cardiac imaging. Motion detection proved valuable for high-resolution imaging, particularly in brain, prostate, and rectal areas. While ACS technology allowed for single-shot fast imaging, it could sometimes sacrifice detail in high-resolution scans. Deep Recon technology automatically enhanced critical textures in MR images, though it sometimes overemphasized background noise. DeepRecon reconstruction significantly improved PET/MR signal-to-noise ratio (SNR) and contrast. Additionally, lesion detection technology showed potential in tumor staging, follow-up, and treatment evaluation, while sCT-based μ-map generation clearly visualized bone structures, notably in rib metastases.

PET/MR systems ensure compatibility between MR and PET subsystems through technologies like SiPM, temperature control, and RF shielding. However, challenges remain, such as the impact of B0 field DSV increases, interference from MR RF gradients, and eddy current heating.¹⁹ Surface coils on the body, which cannot undergo attenuation correction, generate scattering that affects PET image quality. Low-density coils minimize this effect but compromise MR image quality. While AI-assisted imaging showed no significant differences in clinical PET/MR images compared to conventional PET/CT or MR, it greatly improved quantitative accuracy, expanding PET/MR's innovative applications.

Current AI-assisted imaging research often focuses on single technologies’ independent effects, while comprehensive evaluation of combined technologies is lacking. Our study explored the impact of technology fusion on clinical values of PET/MR. We found that the integration of multiple AI technologies significantly shortened MR and PET imaging durations, allowing more time for regional PET imaging, full-body DWI, and high-resolution MR. This shifted the focus from acquisition efficiency to precise PET/MR image alignment, using multi-excitation DWI and Radial T1 W sequences to avoid mismatches from breath-holding.

In conclusion, we quantitatively and qualitatively validated key technologies affecting whole-body PET/MR imaging. The study revealed that AI-assisted technologies optimized PET and MR image quality, with DeepRecon reconstruction enhancing PET image quality and contrast, sCT-based attenuation correction improving PET quantification, and MR ACS and DeepRecon streamlining MR acquisition. The SuperFlex Coil may have significant applications in specialized PET/MR imaging, such as pediatric, breast, prostate, and cardiac cases.

Footnotes

Acknowledgements

We would like to thank the staff involved in data collection for this project.

ORCID iDs

Hanzhong Wang

Zheng Zhang

Qiu Huang

Ethics approval

All procedures involving human participants were conducted in accordance with the ethical guidelines of the 1964 Helsinki Declaration and national regulations. The study was approved by the Research Ethics Committee of Ruijin Hospital affiliated to Shanghai Jiao Tong University School of Medicine.

Informed consent

Informed consent was obtained from all individual participants included in the study.

Consent to publish

All authors have reviewed the final version of the manuscript and approved it for submission to this journal.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was funded by Shanghai Municipal Key Clinical Specialty (shslczdzk03403), Guangci Clinical Technology and Innovation Program (GCTIP) of Ruijin Hospital (YW20220006).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

The datasets generated during and/or analysed during the current study are available from the corresponding author on reasonable request.*

References

Disselhorst

Bezrukov

Kolb

, et al. Principles of PET/MR imaging. J Nucl Med 2014; 55: 2S–10S.

Wang

Huang

Zhao

, et al. [(68)Ga]ga-FAPI-04 PET/MR imaging strategy in management of Krukenberg tumors (KTs) from gastric signet-ring-cell carcinoma: to overcome limitation of [(68)Ga]ga-FAPI-04 PET imaging in KTs. Eur J Nucl Med Mol Imaging 2024; 51: 3440–3449.

Ding

Zhao

Chen

, et al. The Added Value of Contrast-Enhanced 68Ga-FAPI-04 PET/MRI Over 18F-FDG PET/CT in Hepatocellular Carcinoma. Clin Nucl Med 2024; 49: 1039–1040.

Leithner

Horvat

Bernard-Davila

, et al. A multiparametric [(18)F]FDG PET/MRI diagnostic model including imaging biomarkers of the tumor and contralateral healthy breast tissue aids breast cancer diagnosis. Eur J Nucl Med Mol Imaging 2019; 46: 1878–1888.

Henriksen

Larsen

Muhic

, et al. Simultaneous evaluation of brain tumour metabolism, structure and blood volume using [(18)F]-fluoroethyltyrosine (FET) PET/MRI: feasibility, agreement and initial experience. Eur J Nucl Med Mol Imaging 2016; 43: 103–112.

Catalano

Masch

Catana

, et al. An overview of PET/MR, focused on clinical applications. Abdom Radiol (NY) 2017; 42: 631–644.

Kirchner

Deuschl

Grueneisen

, et al.

(18)F-FDG PET/MRI in patients suffering from lymphoma: how much MRI information is really needed?

Eur J Nucl Med Mol Imaging 2017; 44: 1005–1013.

Chen

, et al. Impact of patient comfort on diagnostic image quality during PET/MR exam: a quantitative survey study for clinical workflow management. J Appl Clin Med Phys 2019; 20: 184–192.

Meng

Guan

, et al. Clinical evaluation of MR-gated respiratory motion correction in simultaneous PET/MRI. Clin Nucl Med 2021; 46: 297–302.

10.

Liu

Deng

Zeng

, et al. AI-assisted compressed sensing and parallel imaging sequences for MRI of patients with nasopharyngeal carcinoma: comparison of their capabilities in terms of examination time and image quality. Eur Radiol 2023; 33: 7686–7696.

11.

Kustner

Liebgott

Mauch

, et al. Automated reference-free detection of motion artifacts in magnetic resonance images. MAGMA 2018; 31: 243–256.

12.

Chen

E-J

Safwan Selvam

Lee

, et al. Quantitative analysis evaluation of image reconstruction algorithms between digital and analog PET-CT. Radiation Physics and Chemistry 2024/03/01/; 216: 111401.

13.

Wang

Qiao

Wang

, et al. Deep progressive learning achieves whole-body low-dose (18)F-FDG PET imaging. EJNMMI Phys 2022; 9: 82.

14.

Zhang

Huang

Zhang

, et al. Whole-body lesion segmentation in 18f-fdg pet/ct. arXiv preprint arXiv:2209.07851 2022.

15.

Ahangari

Beck Olin

Kinggård Federspiel

, et al. A deep learning-based whole-body solution for PET/MRI attenuation correction. EJNMMI Phys 2022; 9: 55.

16.

Deller

Mathew

Hurley

, et al. PET Image quality improvement for simultaneous PET/MRI with a lightweight MRI surface coil. Radiology 2021; 298: 166–172.

17.

Wang

Zhao

Xing

, et al. Feasibility of AI-assisted compressed sensing protocols in knee MR imaging: a prospective multi-reader study. Eur Radiol 2023; 33: 8585–8596.

18.

. PET Image reconstruction with deep progressive learning. Phys Med Biol 2021; 66: 10.

19.

Levin

Maramraju

Khalighi

, et al. Design Features and Mutual Compatibility Studies of the Time-of-Flight PET Capable GE SIGNA PET/MR System. IEEE Trans Med Imaging 2016; 35: 1907–1914.