Artificial Intelligence for Radiation Dose Optimization in Pediatric Radiology: A Systematic Review

Radiation dose optimization is particularly important in pediatric radiology, as children are more susceptible to potential harmful effects of ionizing radiation. However, only one narrative review about artificial intelligence (AI) for dose optimization in pediatric computed tomography (CT) has been published yet. The purpose of this systematic review is to answer the question “What are the AI techniques and architectures introduced in pediatric radiology for dose optimization, their specific application areas, and performances?” Literature search with use of electronic databases was conducted on 3 June 2022. Sixteen articles that met selection criteria were included. The included studies showed deep convolutional neural network (CNN) was the most common AI technique and architecture used for dose optimization in pediatric radiology. All but three included studies evaluated AI performance in dose optimization of abdomen, chest, head, neck, and pelvis CT; CT angiography; and dual-energy CT through deep learning image reconstruction. Most studies demonstrated that AI could reduce radiation dose by 36–70% without losing diagnostic information. Despite the dominance of commercially available AI models based on deep CNN with promising outcomes, homegrown models could provide comparable performances. Future exploration of AI value for dose optimization in pediatric radiology is necessary due to small sample sizes and narrow scopes (only three modalities, CT, positron emission tomography/magnetic resonance imaging and mobile radiography, and not all examination types covered) of existing studies.


Introduction
Radiology is an indispensable part of modern healthcare. However, most of the medical imaging modalities, such as computed tomography (CT), positron emission tomography (PET), and general radiography, use ionizing radiation for image production [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16]. Although the radiation dose involved in these imaging modalities is low (<100 mSv), and their real risk is unclear, some epidemiologic and biologic studies have demonstrated that these radiological examinations can cause cancers [17][18][19][20][21][22][23]. Hence, "as low as reasonably achievable" (ALARA) has become the fundamental principle of radiology practice [17,24,25]. International Commission on Radiological Protection (ICRP) has introduced the diagnostic reference levels (DRLs) initiative for radiological departments to identify examinations with radiation doses exceeding their corresponding DRLs and trigger the radiation doseoptimization process [26][27][28][29][30][31][32]. As the radiation used in radiological examinations is the source of signal, a reduction of the radiation amount results in a decrease of signal strength and an increase of image noise. Traditionally, the dose-optimization process involves the manipulation of a range of exposure/scan parameters and identification of parameters that deliver the lowest radiation dose but still producing images able to meet minimal diagnostic

Literature Search
The literature search with use of electronic scholarly publication databases, including Google Scholar, PubMed/Medline, ScienceDirect, Scopus, and Web of Science was conducted on 3 June 2022 to identify articles about the AI for dose optimization in pediatric radiology published between 2017 and 2022. The search statement used was ("Artificial Intelligence" OR "Machine Learning" OR "Deep Learning") AND ("Dose Optimization" OR "Dose Reduction") AND ("Pediatric" OR "Children") AND ("Radiology" OR "Medical Imaging"). The keywords used in the search were based on the review focus. The year range was determined based on a narrative review about current and future applications of AI in radiology, which showed the use of AI for dose optimization in radiology not evident before 2017 [43].

Article Selection
A reviewer with more than 20 years of experience in conducting literature review was involved in the article selection process [44]. Only peer-reviewed original research articles that were written in English and focused on the use of AI for dose optimization in pediatric radiology were included. Grey literature, conference abstracts, editorials, review, perspective, opinion, commentary, and non-peer-reviewed (e.g., those published via the arXiv research-sharing platform, etc.) articles were excluded because of the following reasons: Well-established methodological guidelines were not available for proper selection of grey literature. Conference abstracts could not provide complete study information. Only secondary information was presented in editorials, review, perspective, opinion, and commentary articles. Non-peer-reviewed articles might provide unsubstantiated information [45,46]. Figure 1 illustrates details of the article selection process [41]. A three-stage screening process through assessing (1) article titles, (2) abstracts, and (3) full texts against the selection criteria was employed after duplicate article removal from results of the database search. Every non-duplicate article within the search results was retained until its exclusion could Figure 1 illustrates details of the article selection process [41]. A three-stage screening process through assessing (1) article titles, (2) abstracts, and (3) full texts against the selection criteria was employed after duplicate article removal from results of the database search. Every non-duplicate article within the search results was retained until its exclusion could be decided. Lists of references of the included papers were reviewed for additional, relevant article identification [46].

Data Extraction and Synthesis
A data extraction form (Table 1) was developed based on a recent systematic review on the use of AI in radiology [45]. The data, including names and countries of authors, publication years, clinical domains (radiology/nuclear medicine), AI techniques (such as machine learning and deep learning (DL)), model architectures (e.g., convolutional neural network (CNN), generative adversarial network (GAN), etc.), specific application areas (i.e., examination types and approaches that AI was used to achieve dose optimization), imaging modalities, details of AI model development (i.e., whether homegrown or commercially available model and arrangement of model training and testing), AI model evaluation approach (e.g., phantom study, clinical study, etc.), and key findings of AI model performance in dose optimization (including figures of dose reduction and diagnostic values and subjective and objective image assessment scores), were extracted from each included paper. To facilitate comparison of the AI model performance, percentage of dose reduction (if not reported) was synthesized based on the available absolute dose figures. When multiple image-quality-related figures were reported in a study, the most clinically relevant figures were presented. Diagnostic values were considered the most clinically relevant performance figures, while the objective image assessment scores were determined least relevant [47,48]. Quality assessment scores were determined for all included articles based on the quality assessment tool for studies with diverse designs (QATSDD)

Data Extraction and Synthesis
A data extraction form (Table 1) was developed based on a recent systematic review on the use of AI in radiology [45]. The data, including names and countries of authors, publication years, clinical domains (radiology/nuclear medicine), AI techniques (such as machine learning and deep learning (DL)), model architectures (e.g., convolutional neural network (CNN), generative adversarial network (GAN), etc.), specific application areas (i.e., examination types and approaches that AI was used to achieve dose optimization), imaging modalities, details of AI model development (i.e., whether homegrown or commercially available model and arrangement of model training and testing), AI model evaluation approach (e.g., phantom study, clinical study, etc.), and key findings of AI model performance in dose optimization (including figures of dose reduction and diagnostic values and subjective and objective image assessment scores), were extracted from each included paper. To facilitate comparison of the AI model performance, percentage of dose reduction (if not reported) was synthesized based on the available absolute dose figures. When multiple image-quality-related figures were reported in a study, the most clinically relevant figures were presented. Diagnostic values were considered the most clinically relevant performance figures, while the objective image assessment scores were determined least relevant [47,48]. Quality assessment scores were determined for all included articles based on the quality assessment tool for studies with diverse designs (QATSDD) and expressed as percentages [49]. Less than 50%, 50-70%, and greater than 70% were considered low, moderate, and high study quality, respectively [46].   Expert readers' agreements of tumor diagnosis between standard and AI-processed 6.25% ultra-low-dose PET images (kappa = 0.942 (USA datasets) and 0.912 (Germany datasets)) were significantly greater than the agreements between standard and 6.25% ultra-low-dose PET images (kappa = 0.650 (USA datasets) and 0.834 (Germany datasets)). Diagnostic accuracy of AI-processed 6.25% ultra-low-dose PET images was adequate, representing 93.75% dose reduction capability. When compared with ASiR-V 50%, medium-and high-strength DLIR images of contrast enhanced abdominal (n = 23) and non-contrast (n = 16) and contrast enhanced (n = 12) chest CT had statistically significantly higher subjective image quality score and lower noise (p < 0.001), illustrating its dose reduction potential. When compared with ASiR-V 70%, high-strength DLIR achieved about 70% and 60% dose reductions for pediatric non-contrast abdominal (n = 10) and chest (n = 10) CT, respectively. However, high-strength DLIR did not statistically significantly improve subjective image assessment score of chest CT.

Discussion
The findings of this systematic review on the AI for radiation dose optimization in pediatric radiology are consistent with several recent narrative reviews about the use of AI in radiology [17,43,51]. For the narrative review about the current and future applications of AI in radiology published in 2018 [43], only one study regarding low-dose CT denoising published in 2017 was cited [52]. However, recently, more studies about the use of AI for dose optimization have been published, resulting in a narrative review about the AI for dose optimization in pediatric CT available in 2021 [17]. This demonstrates that the use of AI for dose optimization in pediatric radiology has attracted the attention of the profession recently. That narrative review indicated the DLIR allowed 30-80% dose reduction in pediatric CT but was still able to produce images with diagnostic quality. This systematic review with inclusion of more studies about dose optimization in pediatric CT and covering other imaging modalities shows that the majority of the AI models were able to reduce the radiation dose by 36-70% [1,6,7,10,13,16]. Nonetheless, three studies included in this systematic review demonstrated that the use of AI could achieve further radiation dose reduction (up to 95%) [2,12,14]. Apparently, the large variation of dose reduction performances is due to the retrospective nature of many included studies [1,[4][5][6][7][8][9]12,15], which did not allow further manipulation of examination/scan parameters to obtain ultra-low-dose images for evaluating whether the AI models could restore the quality of these ultra-low-dose images to close to the original [9]. Although there is a greater flexibility for phantom studies to manipulate the examination/scan parameters without any ethical and radiation dose concerns, enabling further exploration of the potential of these AI models, their evaluation outcomes tend to be less clinically relevant [47,48]. For example, Jeon et al. [2] reported that Canon AiCE was able to reduce the CT dose by 95% with the contrastto-noise ratio values of the DLIR phantom images similar to those reconstructed by filtered back projection, but it is unclear whether these findings could be translated into clinical practice exactly. Nonetheless, Wang et al.'s [14] clinical prospective study showed that their homegrown AI denoising model developed through transfer learning with the use of 17 standard-dose PET simulated 6.25% ultra-low-dose PET, and MRI training datasets were able to reduce the radiation dose by 93.8% for the whole-body PET examinations with adequate diagnostic accuracy. This implies that it is feasible to use the AI denoising to achieve about 90% dose reduction in the clinical practice although all included studies had small sample sizes and/or number of training datasets [1,[4][5][6][7][8][9][10][11][12][13][14][15][16], which is a common issue of AI studies in radiology due to limited availability of medical images [53]. Nevertheless, through the use of transfer learning (i.e., retraining an existing AI model using a smaller number of datasets with or without modification of its architecture) to develop an AI model for performing a similar task, such a model could provide a dose-optimization performance comparable to commercially available models (e.g., Canon AiCE, GE TrueFidelity, etc.) trained with more datasets [2,12,14,43].
It is within expectation that all but two studies used the AI models with the deep CNN architecture because the CNN architecture emerged in 1980s, and hence, it has been widely used in radiology, with satisfactory performances well-demonstrated [1,2,[4][5][6][8][9][10][11][12][13][14][15][16]37]. However, one included study published in 2022 employed the more recent and advanced DL architecture: GAN, which was designed in 2014 [7,51]. According to a narrative review about the use of GAN in radiology published in 2021 [51], the CNN-based denoising models could cause CT images having a plastic-like appearance, which is similar to those produced by iterative reconstruction due to over-smoothing. In contrast, the GAN is a more complex architecture with a generator and a discriminator, which requires simultaneous training of these two, increasing the complexity of model development [37]. Nonetheless, the GANbased denoising models are able to preserve texture details and hence produce images with quality matching standard images [51]. The GAN-based dose-optimization study included in this systematic review also demonstrated similar findings that their readers were unable to differentiate between the standard-dose and GAN-processed images although only 36.6% dose reduction was achieved in their study [7]. Another non-CNN-based doseoptimization study included in this review employed the Gaussian mixture model (GMM) architecture [3]. The use of GMM for medical image denoising was reported before the emergence of GAN [54]. However, it is not widely adopted in radiology, and its clinical performance in pediatric radiology dose optimization remains unclear [3,17,43,51]. This paper is the first systematic review on the AI for radiation dose optimization in pediatric radiology covering the imaging modalities of CT, PET/MRI, and mobile radiography and hence advancing the previous narrative review on the AI for dose optimization in pediatric CT published in 2021 [17]. Although it is well-known that radiation dose burden is a significant issue in pediatric CT [1][2][3][4][5][6][7][8][9][10][11]15,16], the dose involved in a PET scan is comparable to that of a CT examination [14]. Furthermore, general radiography is the most common radiological examination type for pediatric patients despite being a lowdose modality [36]. Nonetheless, as per the ALARA principle, the value of AI for dose optimization in other modalities that use ionizing radiation for pediatric examinations should be explored in the future [17,24,25]. Moreover, given the relatively narrow focus and small sample size of the included studies, future studies on this topic area for CT, PET, and general radiography need to have greater scale and wider scope [1,[4][5][6][7][8][9][10][11][12][13][14][15][16]. Besides, further exploration of the use of GAN for dose optimization appears warranted [7,51].
This systematic review has two major limitations. Article selection, data extraction, and synthesis were performed by a single author, albeit one with more than 20 years of experience in conducting literature reviews. According to a recent methodological systematic review [44], this is an appropriate arrangement provided that the single reviewer is experienced. Additionally, through adherence to the PRISMA guidelines and the use of the data extraction form (Table 1) devised based on the recent systematic review on AI in radiology and QATSDD, the potential bias should be addressed to certain extent [41,45,46,49]. In addition, only articles written in English and published within last five years were included, potentially affecting comprehensiveness of this systematic review. Nevertheless, this review still has a wider coverage than the previous narrative review on the AI for dose optimization in pediatric CT [17].

Conclusions
This systematic review shows that the deep CNN was the most common AI technique and architecture used for radiation dose optimization in pediatric radiology. All but three included studies evaluated the AI performance in dose optimization of abdomen, chest, head, neck, and pelvis CT; CT angiography; and DECT through DLIR. The majority of studies demonstrated that the AI could reduce radiation dose by 36-70% without losing diagnostic information. Despite the dominance of commercially available AI models based on the deep CNN, the homegrown models, including the one with the more recent and advanced architecture, i.e., GAN, could provide comparable performances. Future exploration of the value of AI for dose optimization in pediatric radiology is necessary, as the sample sizes of the included studies appear small, and only three imaging modalities, namely CT, PET/MRI, and mobile radiography, rather than all examination types were covered.