Radiomics-Based Machine Learning for Determining MYCN Amplification Status in Childhood Neuroblastoma: A Systematic Review and Meta-Analysis

Abstract

Introduction

The MYCN oncogene promotes tumor cell proliferation in neuroblastoma, and its amplification is a well-established marker of poor prognosis. Radiomics-based approaches have shown promise in noninvasively determining MYCN amplification status; however, their diagnostic performance has varied significantly across studies. This systematic review and meta-analysis aimed to quantitatively evaluate the diagnostic accuracy of radiomics-based machine learning models for determining MYCN amplification in neuroblastoma and to critically assess the methodological quality of the included studies.

Methods

A systematic search of articles published between January 1, 2000, and June 30, 2024, was conducted across PubMed, Embase, Web of Science, and the Cochrane Library. The articles focused on using radiomics to determine MYCN amplification in neuroblastoma. Methodological quality was assessed using the Radiomics Quality Score (RQS), METhodological RadiomICs Score (METRICS), and Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) tools. A meta-analysis of validation performance was performed on studies with Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis statement Type 2a or higher.

Results

Nine studies with 851 patients were included, and seven studies with 217 patients in the validation set were eligible for meta-analysis. The RQS scores ranged from 10 to 16 (mean 12), and METRICS scores ranged from 28.8% to 78.4% (mean 59.7%). QUADAS-2 assessment indicated that most studies had a low or unclear risk of bias. The pooled sensitivity, specificity, positive likelihood ratio, and negative likelihood ratio were 0.78, 0.92, 9.45, and 0.24, respectively. The area under the summary receiver operating characteristic curve was 0.94 (95% confidence interval: 0.91-0.95).

Conclusion

Despite variability in study design and bias risk, radiomics shows promise as a non-invasive method for detecting MYCN amplification in neuroblastoma. Further refinement and validation in multicenter studies with larger sample sizes are needed to enhance its clinical applicability.

Keywords

meta-analysis neuroblastoma radiomics systematic review

Introduction

Neuroblastoma arises from immature neurons of the sympathetic nervous system and often occurs in childhood.^1,2 The MYCN oncogene plays an important role in neuroblastoma pathogenesis, as it can drive both the proliferation and differentiation of tumor cells.³ Amplification of the MYCN oncogene, present in approximately 20% of cases, is a well-established biomarker of high-risk status and is associated with aggressive tumor behavior, rapid progression, and poor prognosis in neuroblastoma.⁴ Regardless of histological subtypes or clinical presentations, neuroblastoma patients with MYCN amplification are classified as high-risk according to the International Neuroblastoma Risk Group (INRG) classification system.⁵ Consequently, assessing MYCN amplification status is crucial for tailoring treatment strategies for neuroblastoma patients. However, accurately determining MYCN amplification status can be challenging due to tumor heterogeneity and potential sampling errors during invasive biopsy.⁶ These limitations highlight the need for non-invasive diagnostic approaches to improve the detection and management of MYCN amplification in neuroblastoma.⁷

Some reports have demonstrated the potential of radiomics-based machine learning in determining the MYCN amplification status in neuroblastoma.^8-10 Radiomics, an emerging technology in medical image analysis, provides a more detailed characterization of tumors than traditional imaging methods by extracting a large number of quantitative voxel features from medical images, thereby enhancing the diagnostic capabilities of current imaging modalities.^11,12 Radiogenomics, a subset of radiomics, links the imaging features with genetic and molecular profiles of tumors, offering a deeper understanding of tumor biology.¹³ This approach allows for non-invasive prediction of genetic alterations, such as MYCN amplification, by analyzing the imaging data. Consequently, identifying MYCN amplification status through radiomics-based machine learning could lead to improve the accuracy of neuroblastoma non-invasive diagnosis.

Despite the potential of radiomics in determining MYCN amplification status in neuroblastoma as shown in multiple studies, its diagnostic performance varies significantly across different reports.^14-16 Some studies have reported that the area under the receiver operating characteristic curve (AUC) for predicting MYCN amplification in neuroblastoma can exceed 0.90,^8,10 while others reported an AUC of only 0.75.¹⁷ This discrepancy may stem from differences in sample size, imaging modalities, feature extraction methods, machine learning algorithms, and study design. Therefore, a systematic evaluation of radiomics studies in determining MYCN amplification status in neuroblastoma is necessary to determine its practical value and limitations. Such an evaluation can lay the foundation for future research, further improving diagnostic performance and clinical applicability of radiomics in determining MYCN amplification status in neuroblastoma.

Therefore, this systematic review and meta-analysis aimed to investigate the effectiveness of radiomics-based machine learning in determining MYCN amplification status in childhood neuroblastoma. Additionally, it aimed to assess the methodological quality of radiomics studies in this context.

Materials and Methods

Registry Information and Evidence Collection Protocol

The study was prospectively registered with the International Prospective Register of Systematic Reviews (PROSPERO) (ID: CRD42024566020) and adhered to the guidelines outlined by the Preferred Reporting Items for Systematic Reviews and Meta-analyses of Diagnostic Test Accuracy Studies (PRISMA-DTA). Since this research involved a secondary analysis of pre-existing literature, ethical approval and informed consent were not required.

Systematic Literature Search and Study Selection Procedure

Two reviewers, each with 4 and 15 years of experience in pediatric imaging, respectively, jointly conducted a search for original articles exploring the application of radiomics in assessing MYCN amplification status in neuroblastoma. The search was performed across PubMed, Embase, Web of Science, and the Cochrane Library for studies published from January 1, 2000, to June 30, 2024. Discrepancies were resolved through discussion to achieve consensus. The search strategy employed was: ((radiomics) OR (radiogenomics) OR (texture analysis) OR (artificial intelligence) OR (deep learning) OR (machine learning)) AND ((neuroblastoma) OR (neuroblastic tumor) OR (NB)) AND ((MYCN) OR (N-Myc)). To ensure thoroughness, no restrictions were placed on imaging modalities during the initial search; however, these modalities were limited during the subsequent manual review. Articles were selected based on the following criteria. Inclusion criteria: (1) Studies published in English with full-text availability; (2) Studies involving patients with neuroblastoma; (3) Use of imaging modalities such as CT, MRI, or PET/CT; (4) Employment of radiomics or texture features from medical images to determine MYCN amplification status. Exclusion criteria: (1) Duplicate studies; (2) Case reports, conference abstracts, reviews, or editorials; (3) Studies focusing exclusively on segmentation or feature extraction algorithms.

Data Extraction

Two reviewers, working together, collected the following information from each study: 1) Study general characteristics: including the authors, year of publication, type and design of the study, and the country of origin; 2) Participant characteristics: total number of cases, cases with and without MYCN amplification, tumor location, INRG stages, and reference standard; 3) Radiomics and machine learning characteristics: imaging modality, imaging phase/sequence, tumor segmentation method, number of extracted and selected radiomics features, modeling and validation approach, as well as sensitivity, specificity, and AUC values. From these details, we computed the numbers for True Positives (TPs), False Positives (FPs), False Negatives (FNs), and True Negatives (TNs). We only documented the performance metrics from the validation set of the proposed radiomics models to avoid overfitting in the training process, so only the studies with Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis (TRIPOD) statement Type 2a or above were included into subsequent meta-analysis. In cases where multiple radiomics models were developed for the same patient group in a study, only the model demonstrating the highest diagnostic performance was included in the meta-analysis.

According to the TRIPOD statement, studies can be classified based on the model's validation method and study design.¹⁸ TRIPOD types include Type 1a (model development and evaluation on the same dataset), Type 1b (internal validation using methods like cross-validation or bootstrapping), Type 2a (validation on a randomly splitting dataset), Type 2b (validation on a non-randomly splitting dataset), Type 3 (validation on a completely independent dataset), and Type 4 (validation of an external model using independent data). Additionally, based on sample size and design, studies can be categorized into Discovery Science (experimental research), Phase 0 (retrospective, internal validation with <100 patients or >100 patients), Phase I (retrospective, external validation with <100 patients), Phase II (retrospective, external validation with >100 patients), Phase III (prospective, external validation with >100 patients), and Phase IV (real-world application).¹⁹

Assessment of Methodological Quality

Two reviewers assessed the methodological quality of the included articles using the Radiomics Quality Score (RQS), METhodological RadiomICs Score (METRICS), and the Quality Assessment of Diagnostic Accuracy Studies (QUADAS-2) tools. The METRICS scores were calculated using a user-friendly online calculation tool https://metricsscore.github.io/metrics/METRICS.html.

RQS is a standardized tool developed by Lambin et al²⁰ to assess the methodological quality of radiomics studies. It consists of 16 scoring items that evaluate key aspects such as the quality of imaging acquisition protocols, image segmentation, feature selection, model validation methods, and clinical utility. The maximum RQS score is 36, which corresponds to 100% on the scoring scale. Under the RQS framework, the total score for a radiomics study can range from −8 to 36, with scores between −8 and 0 defined as 0%. A higher RQS score indicates better methodological rigor and reliability, reflecting higher standards in study design and reporting.

METRICS is a new evaluation framework released in 2024 by a group of international experts, designed to assess the quality and transparency of studies in radiomics and machine learning.²¹ This framework aims to provide detailed guidelines to ensure the reproducibility and reliability of research. METRICS consists of 9 major categories, including study design, imaging data, image segmentation, image processing and feature extraction, model development and validation, results reporting, model performance interpretation, etc. Within these 9 categories, there are 30 specific items, each with corresponding scoring criteria.

QUADAS-2 is a widely used tool for assessing the quality of diagnostic accuracy studies.²² It helps evaluate both the risk of bias and the applicability of studies, ensuring the reliability and clinical relevance of diagnostic research. QUADAS-2 consists of four domains: Patient Selection, Index Test, Reference Standard, and Flow and Timing. Each domain includes signaling questions that aid in evaluating potential bias, with each domain rated as low, high, or unclear risk of bias.

Statistical Analysis

STATA (version 14.0), Review Manager (version 5.4.1) and RStudio (version 4.1.3) were used for statistical analysis. First, Egger's test were applied to assess potential publication bias. We then calculated the pooled sensitivity, specificity, positive likelihood ratio (PLR), and negative likelihood ratio (NLR), diagnostic score, and odds ratio to evaluate the diagnostic performance of the included studies. Based on these metrics, we further plotted the summary receiver operating characteristic (SROC) curve and assessed the diagnostic ability of these studies for determining MYCN amplification status by calculating the AUC and its 95% confidence intervals (CIs). Additionally, we created forest plots and conducted a quantitative analysis of the heterogeneity between studies using the I² test. Significant heterogeneity was considered when the I² value was greater than 50%. To further explore the potential sources of heterogeneity, we performed meta-regression and subgroup analysis to identify factors that may influence the heterogeneity. When the P-value was less than 0.05, it indicated statistical significance.

Results

Selection Process and Overview of Included Studies

A total of nine studies involving 851 patients were included in this systematic review^{8-10,14-17,23,24} (Figure 1). Of these, six studies used CT,^{8-10,14,15,24} two used PET/CT,^16,23 and one used MRI.¹⁷ The general characteristics of the studies included in the review are summarized in Table 1. Among the studies, only one employed deep learning for tumor lesion segmentation,²⁴ while the remaining eight used manual segmentation. All studies adopted a validation approach, with one study using external validation and the other eight relying on internal validation. Table 2 provides an overview of the radiomics and machine learning characteristics of the included studies. Eight studies provided detailed information on the categories of radiomics features used to determine MYCN amplification status (Table 3). According to the TRIPOD statement, two studies were classified as Type 1b,^15,24 five as Type 2a,^{8,10,14,16,23} one as Type 2b,⁹ and one as Type 3.¹⁷ Consequently, seven studies,^{8-10,14,16,17,23} with a total of 217 patients in the validation set, were eligible for subsequent meta-analysis.

Figure 1.

Flowchart for Literature Screening.

Table 1.

General Characteristics of Studies Included in the Systematic Review.

Authors	Year	Study Design	Country	Image Modality	Number of Cases (MYCN+/MYCN-)	Tumor Location	INRG Stage	Reference Standard
Tan E et al ¹⁴	2022	Retro, single-center	Singapore	CT	54 (16/38)	Cervicothoracic: 3 Thoracic: 9 Thoracoabdominal: 4 Abdominopelvic: 38	L1: 7 L2: 15 M: 28 MS: 4	FISH
Chen X et al ¹⁰	2021	Retro, single-center	China	CT	172 (47/125)	Abdomen: 172	N/A	FISH
Liu G et al ¹⁵	2022	Retro, two-center	United States	CT	56 (11/45)	N/A	N/A	Unclear
Qian LD et al ²³	2023	Retro, single-center	China	PET/CT	104 (65/39)	N/A	N/A	FISH
Qian L et al ¹⁶	2022	Retro, single-center	China	PET/CT	122 (20/102)	N/A	N/A	FISH
Wu H et al ⁸	2021	Retro, single-center	China	CT	78 (14/64)	Chest: 7 Abdomen: 56 Pelvis: 8 Junction: 7	L1: 26 L2: 32 M: 16 MS: 4	FISH
Di Giannatale A et al ⁹	2021	Retro, single-center	Italy	CT	99 (29/70)	N/A	L1: 20 L2: 20 M: 55 MS: 4	FISH
Ghosh A et al ¹⁷	2023	Retro, two-center	United States	MRI	120 (28/92)	Adrenal: 55 Extra-adrenal retroperitoneal: 25 Paravertebral thoracic: 15 Paravertebral sacral: 9 Paravertebral lumbar: 5 Paravertebral cervical: 3 Paravertebral (lumbar + thoracic): 5 Mediastinal: 1 Skull base: 1	N/A	FISH
Yeow LY et al ²⁴	2023	Retro, single-center	Singapore	CT	46 (13/33)	Cervicothoracic: 3 Thoracic: 8 Thoracoabdominal: 1 Abdominopelvic: 34	L1: 6 L2: 15 M: 22 MS: 3	FISH

FISH, Fluorescence In Situ Hybridization.

Table 2.

Radiomics and Machine Learning Characteristics of Studies Included in the Systematic Review.

Authors	Image Modality	Image Phase/sequence	Tumor Segmentation Method	Image Preprocessing Method	Extracted Radiomics Features	ICC Analysis	Feature Selection Method	Over Resampling	Selected Radiomics Features	Modeling Methods	Validation	TRIPOD Type	Phase Classifcation
Tan E et al¹⁴	CT	Venous phase	3D manual, 3D Slicer	Voxel discretiztion	107	None	Correlation analysis, mRMR	Yes	2	XGBoost	Yes (internal)	Type 2a	Phase 0
Chen X et al¹⁰	CT	Non-contrast, arterial, and venous phase	3D manual, ITK-SNAP	Voxel resampling	3654	Yes	mRMR, LASSO	None	14	LR, SVM, Bayes, RF	Yes (internal)	Type 2a	Phase 0
Liu G et al¹⁵	CT	Venous phase	3D manual, 3D Slicer	None	105	None	Regularization methods	Yes	Unclear	LR, ER, LASSO regression, RF, SVM, ANN	Yes (internal)	Type 1b	Phase 0
Qian LD et al²³	PET/CT	18F-FDG PET, non-contrast CT	3D manual, 3D Slicer	None	1720	Yes	Mann-Whitney U test, Spearman correlation analysis, LASSO	None	4	LR, DT, SVM	Yes (internal)	Type 2a	Phase 0
Qian L et al¹⁶	PET/CT	18F-FDG PET, non-contrast CT	3D manual, 3D Slicer	None	2632	Yes	LASSO	Yes	6	LR	Yes (internal)	Type 2a	Phase 0
Wu H et al⁸	CT	Non-contrast, arterial, and venous phase	5 largest cross-sectional images, ITK- SNAP	None	1188	Yes	mRMR, LASSO	Yes	7	LR	Yes (internal)	Type 2a	Phase 0
Di Giannatale A et al⁹	CT	Venous phase	2D manual, Eclipse Treatment Planning System	None	232	None	Boruta selection, Pearson correlation analysis	None	2	LR	Yes (internal)	Type 2b	Phase 0
Ghosh A et al¹⁷	MRI	T1WI and T2WI	3D manual, 3D Slicer	Histogram matching, image normalization, voxel resampling	1132	Yes	mRMR	None	50	RF, SVM	Yes (external)	Type 3	Phase II
Yeow LY et al²⁴	CT	Venous phase	Deep learning	ComBat harmonization	42	None	Ensemble approach	Yes	3	LR	Yes (internal)	Type 1b	Phase 0

ICC, intra-class correlation coefficient; TRIPOD, Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis; RQS, radiomics quality score; mRMR, minimum Redundancy Maximum Relevance; LASSO, Least Absolute Shrinkage and Selection Operator; LR, logistic regression; SVM, support vector machine; RF, random forest; DT, decision tree; ANN, artificial neural network.

Table 3.

Summary of the Radiomics Features Used to Determine MYCN Amplification Status.

Authors	Radiomics Feature Categories
Tan E et al ¹⁴	NeighborhoodGrayToneDifferenceMatrix_Strength GrayLevelSizeZoneMatrix_GrayLevelNonUniformity.2
Chen X et al ¹⁰	NP_wavelet.LHL_firstorder_Median VP_original_firstorder_10Percentile AP_log.sigma.5.0.mm.3D_firstorder_Skewness NP_wavelet.LHL_GrayLevelSizeZoneMatrix_ZoneEntropy VP_wavelet.LLH_firstorder_Kurtosis NP_log.sigma.3.0.mm.3D_firstorder_90Percentile VP_wavelet.LLH_GrayLevelDependenceMatrix_DependenceVariance NP_wavelet.HHH_GrayLevelSizeZoneMatrix_ZoneEntropy NP_wavelet.LHL_GrayLevelSizeZoneMatrix_LargeAreaLowGrayLevelEmphasis NP_wavelet.HHH_GrayLevelSizeZoneMatrix_SizeZoneNonUniformity NP_log.sigma.1.0.mm.3D_GrayLevelSizeZoneMatrix_LargeAreaEmphasis VP_wavelet.LHL_GrayLevelSizeZoneMatrix_LargeAreaLowGrayLevelEmphasis VP_log.sigma.1.0.mm.3D_firstorder_Median VP_log.sigma.5.0.mm.3D_GrayLevelCo-occurrenceMatrix_MaximumProbability
Qian LD et al ²³	CT_HLL_NeighborhoodGrayToneDifferenceMatrix_Strength CT_HHL_GrayLevelCo-occurrenceMatrix_IV PET_logarithm_GrayLevelCo-occurrenceMatrix_ClusterShade PET_HHH_GrayLevelSizeZoneMatrix_SALGLE
Qian L et al ¹⁶	PET_squareroot_GrayLevelDependenceMatrix_HighGrayLevelEmphasis PET_wavelet.LHL_GrayLevelDependenceMatrix_DependenceNonUniformity PET_wavelet.HHH_GrayLevelSizeZoneMatrix_SizeZoneNonUniformity CT_logarithm_firstorder_Skewness CT_wavelet.LLL_GrayLevelDependenceMatrix_DependenceVariance CT_wavelet.HLL_GrayLevelSizeZoneMatrix_LargeAreaHighGrayLevelEmphasis
Wu H et al ⁸	VP_Percentile5 VP_RunLengthNonuniformity_AllDirection_offset7_SD AP_ ShortRunEmphasis_AllDirection_offset7_SD NP_RunLengthNonuniformity_AllDirection_offset1_SD AP_ShortRunHighGreyLevelEmphasis_AllDirection_offset1_SD NP_Maximum3DDiameter NP_uniformity
Di Giannatale A et al ⁹	F_stat.mean F_szm_2.5D.zsnu (zone size nonuniformity)
Ghosh A et al ¹⁷	Wavelet.LLL_firstorder_Energy_t1 Log.sigma.3.0.mm.3D_GrayLevelRunLengthMatrix_RunEntropy_t2 Wavelet.LHL_firstorder_Minimum_t2 Wavelet.LHL_firstorder_Maximum_t1 Wavelet.LLH_firstorder_Maximum_t2 original_shape_Maximum2DDiameterSlice_t2 Wavelet.LLL_GrayLevelCo-occurrenceMatrix_JointAverage_t2
Yeow LY et al ²⁴	NeighborhoodGrayToneDifferenceMatrix_Strength NeighborhoodGrayToneDifferenceMatrix_Busyness GrayLevelSizeZoneMatrix_SizeZoneNonUniformity

RQS, METRICS, and Bias Risk Assessment

The RQS scores of the included studies ranged from 10 to 16, with a mean of 12 and a standard deviation of 2. Specifically, two studies had an RQS score of 10,^9,15 two studies scored 12,^14,16 and another two scored 14.^8,10 Detailed RQS scoring information for each included study is presented in Table 4. The METRICS scores of the included studies ranged from 28.8% to 78.4%, with a mean of 59.7% and a standard deviation of 13.8%. Among these, five studies were categorized as having “Good” quality,^{8,10,17,23,24} three were rated as “Moderate”,^9,14,16 and one was rated as “Low”.¹⁵ Table 5 provides detailed METRICS scoring information for each included study. Based on the QUADAS-2 evaluation, most studies showed either a low or unclear risk of bias. The results of the risk of bias and applicability concerns, as assessed using the QUADAS-2 tool, are presented in Figure 2.

Figure 2.

Methodological Quality Assessment of the Included Studies Using the Quality Assessment of Diagnostic Accuracy Studies 2 (QUADAS-2) Tool. (A) Risk of Bias Summary: Reviewers’ Judgments for Each Risk of Bias Item Across All Included Studies. (B) Risk of Bias Graph: Percentage Representation of Reviewers’ Judgments for Each Risk of Bias Item Across the Included Studies.

Table 4.

The Radiomics Quality Score of the Included Studies.

Criteria	Tan E et al ¹⁴	Chen X et al ¹⁰	Liu G et al ¹⁵	Qian LD et al ²³	Qian L et al ¹⁶	Wu H et al ⁸	Di Giannatale A et al ⁹	Ghosh A et al ¹⁷	Yeow LY et al ²⁴
Image protocol quality	+1	+1	0	+1	0	+1	0	+1	+1
Multiple segmentations	0	+1	0	+1	+1	+1	0	+1	0
Phantom study on all scanners	0	+1	0	0	0	0	0	0	0
Imaging at multiple time points	0	0	0	0	0	0	0	0	0
Feature reduction or adjustment for multiple testing	+3	+3	+3	+3	+3	+3	+3	+3	+3
Multivariable analysis with non radiomics features	+1	+1	0	+1	+1	+1	0	0	0
Detect and discuss biological correlates	+1	+1	+1	+1	+1	+1	+1	+1	+1
Cut-off analyses	0	0	0	0	0	0	0	0	0
Discrimination statistics	+1	+1	+2	+2	+2	+1	+2	+1	+1
Calibration statistics	+1	+1	0	+1	0	0	0	0	0
Prospective study registered in a trial database	0	0	0	0	0	0	0	0	0
Validation	+2	+2	+2	+2	+2	+2	+2	+3	+3
Comparison to “gold standard”	+2	+2	+2	+2	+2	+2	+2	+2	+2
Potential clinical utility	0	0	0	+2	0	+2	0	0	0
Cost-effectiveness analysis	0	0	0	0	0	0	0	0	0
Open science and data	0	0	0	0	0	0	0	+1	0
Radiomics quality score (percentage)	12 (12/36, 33.3%)	14 (14/36, 38.9%)	10 (10/36, 27.8%)	16 (16/36, 44.4%)	12 (12/36, 33.3%)	14 (14/36, 38.9%)	10 (10/36, 27.8%)	13 (13/36, 36.1%)	11 (11/36, 30.6%)

Table 5.

The METhodological RadiomICs Score (METRICS) of the Included Studies.

Categories	Items	Weights	Tan E et al ¹⁴	Chen X et al ¹⁰	Liu G et al ¹⁵	Qian LD et al ²³	Qian L et al ¹⁶	Wu H et al ⁸	Di Giannatale A et al ⁹	Ghosh A et al ¹⁷	Yeow LY et al ²⁴
Study design	Adherence to radiomics and/or machine learning-specifc checklists or guidelines	0.0368	no	no	no	no	no	no	no	no	no
	Eligibility criteria that describe a representative study population	0.0735	yes	yes	no	yes	yes	yes	yes	yes	yes
	High-quality reference standard with a clear definition	0.0919	yes	yes	no	yes	yes	yes	yes	yes	yes
Imaging data	Multi-center	0.0438	no	no	no	no	no	no	no	yes	no
	Clinical translatability of the imaging data source for radiomics analysis	0.0292	yes	yes	yes	yes	yes	yes	yes	yes	yes
	Imaging protocol with acquisition parameters	0.0438	yes	yes	no	yes	no	yes	yes	yes	yes
	The interval between imaging used and reference standard	0.0292	no	yes	no	no	no	no	no	no	no
Segmentation	Transparent description of segmentation methodology	0.0337	yes	yes	yes	yes	yes	yes	yes	yes	yes
	Formal evaluation of fully automated segmentation	0.0225	n/a	n/a	n/a	n/a	n/a	n/a	n/a	n/a	yes
	Test set segmentation masks produced by a single reader or automated tool	0.0112	no	no	no	no	yes	no	no	yes	yes
Image processing and fea-ture extraction	Appropriate use of image preprocessing techniques with transparent description	0.0622	no	yes	no	no	no	no	no	yes	yes
	Use of standardized feature extraction software	0.0311	yes	yes	yes	yes	yes	yes	yes	yes	yes
	Transparent reporting of feature extraction parameters, otherwise providing a default confguration statement	0.0415	no	no	no	yes	yes	yes	yes	yes	yes
Feature processing	Removal of non-robust features	0.0200	no	yes	no	yes	yes	yes	no	yes	no
	Removal of redundant features	0.0200	yes	yes	yes	yes	yes	yes	yes	yes	yes
	Appropriateness of dimensionality compared to data size	0.0300	yes	yes	no	yes	yes	yes	yes	yes	yes
	Robustness assessment of end-to-end deep learning pipelines	0.0200	n/a	n/a	no	n/a	n/a	n/a	n/a	n/a	n/a
Preparation for modeling	Proper data partitioning process	0.0599	yes	yes	yes	yes	yes	yes	yes	yes	no
	Handling of confounding factors	0.0300	no	no	no	no	no	no	no	no	no
Metrics and comparison	Use of appropriate performance evaluation metrics for task	0.0352	yes	yes	yes	yes	yes	yes	yes	yes	yes
	Consideration of uncertainty	0.0234	yes	yes	yes	yes	yes	yes	no	yes	no
	Calibration assessment	0.0176	no	yes	no	yes	no	no	no	no	no
	Use of uni-parametric imaging or proof of its inferiority	0.0117	yes	yes	yes	no	no	no	yes	no	yes
	Comparison with a non-radiomic approach or proof of added clinical value	0.0293	no	yes	no	yes	yes	yes	no	no	no
	Comparison with simple or classical statistical models	0.0176	no	no	no	yes	no	no	no	yes	yes
Testing	Internal testing	0.0375	yes	yes	yes	yes	yes	yes	yes	no	yes
	External testing	0.0749	no	no	no	no	no	no	no	yes	no
Open science	Data availability	0.0075	no	no	no	no	no	no	no	no	no
	Code availability	0.0075	no	no	no	no	no	no	no	yes	no
	Model availability	0.0075	no	no	no	no	no	no	no	no	no
Total METRICS score			54.4%	70.9%	28.8%	66.3%	59.3%	62.7%	56.3%	78.4%	60.5%
Quality category			Moderate	Good	Low	Good	Moderate	Good	Moderate	Good	Good

Meta-Analysis

Egger's test indicated no publication bias among the included studies (P = 0.731). The pooled sensitivity, specificity, PLR, NLR, diagnostic score, and odds ratio for the radiomics-based machine learning model determining MYCN amplification status were 0.78 (95% CI: 0.49-0.93), 0.92 (95% CI: 0.81-0.97), 9.45 (95% CI: 4.46-20.00), 0.24 (95% CI: 0.09-0.64), 3.68 (95% CI: 2.58-4.78), and 39.66 (95% CI: 13.16-119.51), respectively. Figure 3 displays the forest plots for these metrics. The I² values for sensitivity, specificity, and NLR across the studies were 77.14%, 49.31%, and 90.77%, respectively, suggesting substantial heterogeneity among the included studies. The AUC of the SROC curve for determining MYCN amplification status was 0.94 (95% CI: 0.91-0.95), indicating high diagnostic performance (Figure 4).

Figure 3.

Forest Plots of Metrics for Radiomics-Based Machine Learning Models Used to Determine MYCN Amplification Status: (A) Sensitivity, (B) Specificity, (C) Positive Likelihood Ratio, (D) Negative Likelihood Ratio, (E) Diagnostic Score, and (F) Odds Ratio.

Figure 4.

Summary Receiver Operating Characteristic (SROC) Curve of Radiomics-Based Machine Learning Models Used to Determine MYCN Amplification Status in Neuroblastoma. AUC, Area Under the Curve.

Meta-Regression and Subgroup Analyses

The meta-regression analysis identified imaging modality as a significant contributor to heterogeneity among studies (P = 0.047) (Figure 5A). While over resampling appeared to influence heterogeneity, its effect did not reach statistical significance (P = 0.061) (Figure 5B). Other factors, including the number of patients (P = 0.116), intra-class correlation coefficient analysis (P = 0.255), and image preprocessing approach (P = 0.266), were found to have no significant impact on heterogeneity. The detailed results of the meta-regression and subgroup analyses across various factors are illustrated in Figure 5.

Figure 5.

Meta-Regression and Subgroup Analyses Evaluating the Influence of Imaging Modality (A), Number of Patients (B), Image Preprocessing (C), ICC Analysis (D), and Over Resampling (E) on Heterogeneity Across Different Studies. in Panel (A), “1”, “2”, and “3” Represent CT, MRI, and PET/CT, Respectively. in Panel (B), “1” and “2” Indicate Patient Numbers ≥100 and <100, Respectively. in Panels (C), (D), and (E), “1” Denotes Studies That Utilized the Corresponding Methods, While “2” Denotes Studies That Did Not. CI, Confidence Interval; ICC, Intra-Class Correlation Coefficient.

Discussion

The determination of MYCN amplification status is crucial for risk stratification in patients with neuroblastoma. In this study, we systematically reviewed nine studies that utilized radiomics to determine MYCN amplification status. Although these studies showed variability in design and sample size, their overall quality was acceptable, with five studies achieving a “Good” quality rating based on the METRICS scoring system. Additionally, we performed a meta-analysis of the performance of radiomics models on validation datasets, revealing an SROC-AUC value of 0.94 (95% CI: 0.91-0.95). These findings suggest that radiomics has a high diagnostic efficacy in identifying MYCN amplification status. However, significant heterogeneity was observed among the published studies, with image modality identified as a major contributing factor. Therefore, future research still needs to emphasize standardization in model development and validation processes to enhance reproducibility and facilitate clinical translation.

The recommended imaging modalities for neuroblastoma typically include ultrasound, CT, MRI, and MIBG.²⁵ Each modality plays a unique role in the clinical management of neuroblastoma. Since neuroblastoma primarily occurs in pediatric patients, efforts should be made to minimize the additional radiation burden from imaging examinations. Therefore, CT scans are often performed using a single-phase venous imaging protocol. Due to its short examination time and the absence of a need for anesthesia, CT remains widely used for preoperative evaluation of thoracoabdominal neuroblastomas. Of the nine studies included, six utilized CT imaging,^{8-10,14,15,24} with four specifically employing the recommended venous phase.^9,14,15,24 Our meta-analysis revealed no significant heterogeneity among studies using CT images, suggesting that future research should prioritize single venous phase imaging for assessing MYCN amplification.

It is noteworthy that image modality image was identified as a major contributing factor for heterogeneity among studies. The findings showed that radiomics models based on CT images performed better in identifying MYCN amplification status compared to MRI. This superiority may be attributed to the higher spatial resolution and sensitivity of CT imaging in capturing subtle differences in tissue density. A previous study also highlighted that venous phase CT imaging radiomics outperforms MRI in the differential diagnosis of pathological neuroblastoma subtypes.²⁶ However, current studies on MRI radiomics in neuroblastoma have primarily focused on conventional T1WI and T2WI images. Whether multiparametric MRI could offer additional insights for determining MYCN amplification remains an area for further exploration. Additionally, two studies employed ¹⁸F-FDG PET/CT, which demonstrated high predictive value for MYCN amplification. However, ¹⁸F-FDG PET/CT is not a standard imaging modality for neuroblastoma evaluation. Given the superior diagnostic accuracy of MIBG for detecting bone metastases in neuroblastoma, PET/CT is less widely used compared to MIBG.

In the present review, eight studies reported detailed radiomics feature categories, including first-order statistics (eg, skewness, kurtosis, percentile values), texture features derived from Gray Level Co-occurrence Matrix, Gray Level Size Zone Matrix, and Gray Level Dependence Matrix, as well as shape-based descriptors.^{8-10,14,16,17,23,24} Among these, texture features predominated in models for identifying MYCN amplification. A CT-based study extracted over 1200 features across triphasic scans, ultimately selecting 14 optimal predictors, including eight wavelet-transformed and five LoG-transformed texture features, demonstrating that higher-order filters enhanced discriminatory performance compared to original and first-order features.¹⁰ Collectively, these findings highlight that texture-rich, higher-order radiomics features are particularly sensitive to the intratumoral heterogeneity characteristic of MYCN amplification in neuroblastoma, especially those derived from Gray Level Size Zone Matrix, wavelet, and LoG filters.

Although the clinical value of radiomics has been demonstrated in various diseases, its low reproducibility remains a pressing issue.²⁷ As a methodological approach, radiomics has prompted the academic community to propose stricter standards for its workflows and methods to maximize its potential for clinical translation. Efforts such as the RQS and METRICS scoring frameworks aim to enhance the overall quality of radiomics research.^20,21 Among the nine studies we included, only four applied image preprocessing, and five conducted ICC analyses. Furthermore, significant variability was observed in the machine learning algorithms used to build models across different studies.

More importantly, MYCN amplification occurs in only approximately 20% of neuroblastoma cases, resulting in a significant class imbalance that may compromise the performance and generalizability of machine learning models. Classifiers trained on such imbalanced datasets often exhibit artificially high overall accuracy or AUC by disproportionately favoring the majority class, thereby obscuring poor sensitivity in detecting the clinically critical MYCN-amplified cases. Although data-level solutions such as oversampling are commonly used to mitigate class imbalance, only five of the reviewed studies^{8,10,16,17,23} reported implementing such techniques. Moreover, oversampling can introduce new challenges, including overfitting to synthetic data and reduced model performance on external validation cohorts. To enhance the reliability and clinical applicability of radiomics models, future studies should predefine strategies for managing class imbalance and prioritize validation using independent, real-world datasets.

Significantly, based on the QUADAS-2 assessment tool, we found that most studies had an unclear rating for the “Flow and Timing” domain. The majority of studies did not specify the time interval between imaging acquisition and MYCN gene testing, nor did they report whether time differences could introduce bias. While FISH testing is regarded as the best method for determining MYCN amplification status,²⁸ the strong heterogeneity of neuroblastoma may be influenced by temporal and spatial factors.⁶ Therefore, future studies should not only use imaging and genetic data obtained prior to treatment but also report the time interval between imaging acquisition and MYCN gene testing. Additionally, since MYCN amplification most commonly occurs in abdominal neuroblastomas, particularly those located in the adrenal glands,²⁹ and neuroblastoma can arise at any location within the peripheral sympathetic nervous system, future research should also assess the generalizability of radiomics models predicting MYCN amplification across different anatomical sites.

However, our study still had several limitations. First, due to the rarity of neuroblastoma, with a low prevalence of MYCN amplification, resulting in a limited number of studies available for inclusion in our review. Despite these constraints, a systematic review is still particularly valuable at this stage, as it provides a timely summary of the current evidence, highlights methodological limitations, and identifies key areas for improvement. Second, all the included studies were retrospective, which may have introduced selection bias in patient inclusion. Lastly, the small number of relevant studies, most of which focused on CT imaging (with only one study utilizing MRI and two studies employing PET/CT to detect MYCN amplification), restricted our ability to perform comprehensive meta-analyses for MRI and PET/CT.

Conclusion

In conclusion, despite variability in study design and potential bias, current evidence still demonstrates that radiomics-based machine learning as a non-invasive method holds promise for detecting MYCN amplification in neuroblastoma. However, the diagnostic performance of radiomics-based machine learning for identifying MYCN amplification may vary across different image modalities. Therefore, further refinement and validation through larger-scale, multicenter studies using recommended image modalities are necessary to enhance its clinical applicability.

Footnotes

Abbreviations

ORCID iD

Haoru Wang

Author Contributions

JC and HW conceived the project. HW, XC, LH, and XF collected the data. HW and YJ analyzed the data and performed meta-analysis. HW wrote the first draft of the manuscript. JC revised this manuscript. All authors read and approved the final manuscript.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

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

All data generated or analyzed during this study are included in this published article.

References

Pudela

Balyasny

Applebaum

. Nervous system: Embryonal tumors: Neuroblastoma. Atlas Genet Cytogenet Oncol Haematol. 2020;24(7):284-290.

Huang

Gong

Deng

, et al. Crosstalk betweencancer cells and the nervous system. Med Adv. 2023;1(3):173-189.

Müller

Trunk

Fleischhauer

, et al. MYCN In neuroblastoma: The kings’ new clothes and drugs. EJC Paediatr Oncol. 2024;4:100182.

Zimmerman

Liu

, et al. MYC Drives a subset of high-risk pediatric neuroblastomas and is activated through mechanisms including enhancer hijacking and focal enhancer amplification. Cancer Discov. 2018;8(3):320-335.

Cohn

Pearson

London

, et al.; INRG Task Force. The International Neuroblastoma Risk Group (INRG) classification system: An INRG Task Force report. J Clin Oncol. 2009;27(2):289-297.

Schmelz

Toedling

Huska

, et al. Spatial and temporal intratumour heterogeneity has potential consequences for single biopsy-based neuroblastoma treatment decisions. Nat Commun. 2021;12(1):6804.

Trigg

Turner

Shaw

, et al. Diagnostic accuracy of circulating-free DNA for the determination of MYCN amplification status in advanced-stage neuroblastoma: A systematic review and meta-analysis. Br J Cancer. 2020;122(7):1077-1084.

Zheng

, et al. Radiogenomics of neuroblastoma in pediatric patients: CT-based radiomics signature in predicting MYCN amplification. Eur Radiol. 2021;31(5):3080-3089.

Di Giannatale

Di Paolo

Curione

, et al. Radiogenomics prediction for MYCN amplification in neuroblastoma: A hypothesis generating study. Pediatr Blood Cancer. 2021;68(9):e29110.

10.

Chen

Wang

Huang

, et al. CT-Based Radiomics signature with machine learning predicts MYCN amplification in pediatric abdominal neuroblastoma. Front Oncol. 2021;11:687884.

11.

Jin

Jiang

Zhao

, et al. Radiomics-based machine learning to predict the recurrence of hepatocellular carcinoma: A systematic review and meta-analysis. Acad Radiol. 2024;31(2):467-479.

12.

Berbís

MÁ

Godino

Rodríguez-Comas

, et al. Radiomics in CT and MR imaging of the liver and pancreas: Tools with potential for clinical application. Abdom Radiol (NY). 2024;49(1):322-340.

13.

Singh

Mohajer

Wells

, et al. Imaging genomics and multiomics: A guide for beginners starting radiomics-based research. Acad Radiol. 2024;31(6):2281-2291.

14.

Tan

Merchant

, et al. CT-based morphologic and radiomics features for the classification of MYCN gene amplification status in pediatric neuroblastoma. Childs Nerv Syst. 2022;38(8):1487-1495.

15.

Liu

Poon

Zapala

, et al. Incorporating radiomics into machine learning models to predict outcomes of neuroblastoma. J Digit Imaging. 2022;35(3):605-612.

16.

Qian

Yang

Zhang

, et al. Prediction of MYCN amplification, 1p and 11q aberrations in pediatric neuroblastoma via Pre-therapy 18F-FDG PET/CT radiomics. Front Med (Lausanne). 2022;9:840777.

17.

Ghosh

Yekeler

Teixeira

, et al. Role of MRI radiomics for the prediction of MYCN amplification in neuroblastomas. Eur Radiol. 2023;33(10):6726-6735.

18.

Moons

Altman

Reitsma

, et al. Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis (TRIPOD): Explanation and elaboration. Ann Intern Med. 2015;162(1):W1-W73.

19.

Huang

Ren

Jin

, et al. A systematic review and meta-analysis of CT and MRI radiomics in ovarian cancer: Methodological issues and clinical utility. Insights Imaging. 2023;14(1):117.

20.

Lambin

Leijenaar

RTH

Deist

, et al. Radiomics: The bridge between medical imaging and personalized medicine. Nat Rev Clin Oncol. 2017;14(12):749-762.

21.

Kocak

Akinci

Antonoli

, et al. METhodological RadiomICs Score (METRICS): A quality scoring tool for radiomics research endorsed by EuSoMII. Insights Imaging. 2024;15(1):8.

22.

Whiting

Rutjes

Westwood

, et al.; QUADAS-2 Group. QUADAS-2: A revised tool for the quality assessment of diagnostic accuracy studies. Ann Intern Med. 2011;155(8):529-536.

23.

Qian

Zhang

, et al. Predicting MYCN amplification in paediatric neuroblastoma: Development and validation of a 18F-FDG PET/CT-based radiomics signature. Insights Imaging. 2023;14(1):205.

24.

Yeow

Teh

, et al. Prediction of MYCN gene amplification in pediatric neuroblastomas: Development of a deep learning-based tool for automatic tumor segmentation and comparative analysis of computed tomography-based radiomics features harmonization. J Comput Assist Tomogr. 2023;47(5):786-795.

25.

Littooij

de Keizer

. Imaging in neuroblastoma. Pediatr Radiol. 2023;53(4):783-787.

26.

Wang

Chen

, et al. A head-to-head comparison of computed tomography- and magnetic resonance imaging-based radiomics in assessing pediatric peripheral neuroblastic tumor cell behavior. Abdom Radiol (NY). 2024;49(8):2942-2952.

27.

Park

Kim

, et al. Reproducibility and generalizability in radiomics modeling: Possible strategies in radiologic and statistical perspectives. Korean J Radiol. 2019;20(7):1124-1137.

28.

Santiago

Tarek

Boulos

, et al. Correlation between MYCN gene Status and MYCN protein expression in neuroblastoma: A pilot study to propose the use of MYCN immunohistochemistry in limited-resource areas. J Glob Oncol. 2019;5:1-7.

29.

Brisse

Blanc

Schleiermacher

, et al. Radiogenomics of neuroblastomas: Relationships between imaging phenotypes, tumor genomic profile and survival. PLoS One. 2017;12(9):e0185190.