Generative Adversarial Network (Generative Artificial Intelligence) in Pediatric Radiology: A Systematic Review

Generative artificial intelligence, especially with regard to the generative adversarial network (GAN), is an important research area in radiology as evidenced by a number of literature reviews on the role of GAN in radiology published in the last few years. However, no review article about GAN in pediatric radiology has been published yet. The purpose of this paper is to systematically review applications of GAN in pediatric radiology, their performances, and methods for their performance evaluation. Electronic databases were used for a literature search on 6 April 2023. Thirty-seven papers met the selection criteria and were included. This review reveals that the GAN can be applied to magnetic resonance imaging, X-ray, computed tomography, ultrasound and positron emission tomography for image translation, segmentation, reconstruction, quality assessment, synthesis and data augmentation, and disease diagnosis. About 80% of the included studies compared their GAN model performances with those of other approaches and indicated that their GAN models outperformed the others by 0.1–158.6%. However, these study findings should be used with caution because of a number of methodological weaknesses. For future GAN studies, more robust methods will be essential for addressing these issues. Otherwise, this would affect the clinical adoption of the GAN-based applications in pediatric radiology and the potential advantages of GAN could not be realized widely.

Pediatric radiology is a subset of radiology [26,28,29,66,67].The aforementioned review findings may not be applicable to pediatric radiology [28,29,[55][56][57][58][59][60][61][62]67].For example, the application of GAN for prostate cancer segmentation appears not relevant to children [60,68].Although several literature reviews about AI in pediatric radiology have been published, none of them focused on the GAN [26,28,29,67].Given that the GAN is an important topic area in radiology and the recent literature reviews focused on its applications in this discipline, it is timely to conduct a systematic review of its applications in pediatric radiology [29,[55][56][57][58][59][60][61][62].The purpose of this article is to systematically review published original studies to answer the question "What are the applications of GAN in pediatric radiology, their performances, and methods for their performance evaluation?".

Materials and Methods
This systematic review of the GAN in pediatric radiology was carried out according to the PRISMA guidelines and patient/population, intervention, comparison, and outcome (PICO) model (Table 1) [26,29,69].Four major processes, literature search, article selection, and data extraction and synthesis were involved [26,29].
Table 1.Patient/population, intervention, comparison, and outcome table for the systematic review of the generative adversarial network (GAN) in pediatric radiology.

Inclusion Criteria
Exclusion Criteria
Written in English 3.
Focused on the use of generative adversarial networks in pediatric radiology Non-peer-reviewed article (e.g., paper on the arXiv platform) The exclusion criteria of Table 2 were established because of: 1. unavailability of well-developed methodological guidelines for appropriate grey literature selection; 2. Incomplete study information given in conference abstracts; 3. a lack of primary evidence in editorials, reviews, perspectives, opinions, and commentary; and 4. unsubstantiated information given in non-peer-reviewed papers [26,29,62,71].The detailed process of the article selection is shown in Figure 1 [26,29,69].Duplicate papers were first removed from the database search results.Subsequently, article titles, abstracts, and full texts were assessed against the selection criteria.Each non-duplicate paper in the search results was kept unless a decision on its exclusion could be made.Additionally, relevant articles were identified by checking reference lists of the included papers [26,29,71].

Inclusion Criteria
Exclusion Criteria
Written in English 3.
Focused on the use of generative adversarial networks in pediatric radiology Non-peer-reviewed article (e.g., paper on the arXiv platform) The exclusion criteria of Table 2 were established because of: 1. unavailability of welldeveloped methodological guidelines for appropriate grey literature selection; 2. Incomplete study information given in conference abstracts; 3. a lack of primary evidence in editorials, reviews, perspectives, opinions, and commentary; and 4. unsubstantiated information given in non-peer-reviewed papers [26,29,62,71].The detailed process of the article selection is shown in Figure 1 [26,29,69].Duplicate papers were first removed from the database search results.Subsequently, article titles, abstracts, and full texts were assessed against the selection criteria.Each non-duplicate paper in the search results was kept unless a decision on its exclusion could be made.Additionally, relevant articles were identified by checking reference lists of the included papers [26,29,71].

Data Extraction and Synthesis
Three systematic reviews on the GAN for image classification and segmentation in radiology [62], AI for radiation dose optimization [26] and CAD in pediatric radiology [29], and one narrative review about the GAN in adult brain imaging [56] were used to develop a data extraction form (Table 3).The data, author name and country, publication year, imaging modality, GAN architecture (such as cycle-consistent GAN (CycleGAN)), study design (either prospective or retrospective), patient/population (e.g., 0-10-year-old children), dataset source (such as public cardiac magnetic resonance imaging (MRI) dataset by Children's Hospital Los Angeles, USA) and size (e.g., total: 33 scans-training: 25; validation: 4; testing: 4, etc.), any sample size calculation, application area (such as image synthesis and data augmentation), model commercial availability, model internal validation type (e.g., 4-fold cross-validation, etc.), any model external validation (i.e., any testing of model based on dataset not used in internal validation and obtained from different setting), reference standard for establishing ground truth (such as expert consensus), any comparison of performance of model with clinician, and key findings of model performance (e.g., area under receiver operating characteristic curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and F1 score, etc.) were extracted from every article included [26,29,56,62].For facilitating GAN model performance comparison, improvement figures such as improvement percentages when the GAN was used were synthesized (if not reported) based on the available absolute figures (if feasible) [26].When a study reported performances for more than one GAN model, only the best-performing model performance values were shown [29,72].Meta-analysis was not performed as this systematic review included a range of GAN applications, resulting in high study heterogeneity which would affect its usefulness [29,[73][74][75].The quality assessment tool for studies with diverse designs (QATSDD) was used to determine quality percentages for all included papers [26,71,76].<50%, 50-70%, and >70% represented low, moderate, and high qualities of study, respectively [26,71].
It is noted that no GAN model of the included studies was commercially available .Again, it is within expectation because the GAN has only emerged for 10 years.In contrast, another common DL architecture in medical imaging, convolutional neural network (CNN) which is a deductive AI technique has been available since the 1980s and hence some commercial companies have already used it for developing various products such as Canon Medical Systems Advanced Intelligent Clear-IQ Engine (AiCE) (Tochigi, Japan), General Electric Healthcare TrueFidelity (Chicago, IL, USA), ClariPI ClariCT.AI (Seoul, Republic of Korea), Samsung Electronics Co., Ltd.SimGrid (Suwon-si, Republic of Korea) and Subtle Medical SubtlePET 1.3 (Menlo Park, CA, USA) for radiation dose optimization (denoising) in pediatric CT, X-ray and PET, respectively [1,26].
As a result of the increasing number of GAN publications in pediatric radiology and the popularity of another generative AI application, Chat Generative Pre-Trained Transformer (ChatGPT), it is expected that the GAN would attract the attention of commercial companies to consider using it to develop various applications in pediatric radiology in the future [54,.However, based on the previous trend of CNN-based commercial product development for pediatric radiology, such GAN-based commercial solutions should not be available in the coming few years [1,26].
There are two major limitations in this systematic review.A single author, despite having experience in performing literature reviews for more than 20 years, selected articles, and extracted and synthesized data [26,29].As per a recent methodological systematic review, this arrangement is appropriate as the single reviewer is experienced [26,29,70,[121][122][123]. Additionally, the potential bias would be addressed to a certain degree due to the use of PRISMA guidelines, data extraction form (Table 3) developed based on the recent systematic reviews on GAN for image classification and segmentation in radiology, and AI for radiation dose optimization and CAD in pediatric radiology, and one narrative review about GAN in adult brain imaging, and QATSDD [26,29,56,62,69,76].In addition, only English papers were included and this could potentially affect the systematic review comprehensiveness [26,29,72,[124][125][126]. Nevertheless, a wider range of applications of GAN in pediatric radiology has been covered in this review when compared with the previous review papers [26,28,29,67].

Conclusions
This systematic review shows that the GAN can be applied to pediatric MRI, X-ray, CT, ultrasound, and PET for image translation, segmentation, reconstruction, quality assessment, synthesis and data augmentation, and disease diagnosis.About 80% of the included studies compared their GAN model performances with those of other approaches and indicated that their GAN models outperformed the others by 0.1-158.6%.Also, the absolute performance figures of the best-performing models appear competitive with the other state-of-the-art approaches.However, these study findings should be used with caution because of a number of methodological weaknesses including no sample size calculation, small dataset size, narrow data variety, limited use of cross-validation, patient cohort coverage and disclosure of reference standards, retrospective data collection, overreliance on public dataset, lack of model external validation and model performance comparison with pediatric clinicians.More robust methods will be necessary in future GAN studies for addressing the aforementioned methodological issues.Otherwise, trustworthy findings for the commercialization of these models could not be obtained.Additionally, this would affect the clinical adoption of the GAN-based applications in pediatric radiology and the potential advantages of GAN would not be realized widely.

Figure 1 .
Figure 1.PRISMA flow diagram for the systematic review of the generative adversarial network (GAN) in pediatric radiology.

Figure 1 .
Figure 1.PRISMA flow diagram for the systematic review of the generative adversarial network (GAN) in pediatric radiology.

Table 2 .
Article inclusion and exclusion criteria.

Table 3 .
Characteristics of generative adversarial network (GAN) studies in pediatric radiology (grouped by their applications).

Table 4 .
Absolute performance figures of best-performing generative adversarial network (GAN) models for various applications in pediatric radiology.