Our institution’s institutional review board approved this retrospective study. From January 2019 to September 2021, we obtained the clinical data of 154 patients with NB confirmed by pathology and collected their preoperative 18F-FDG PET/CT images. The inclusion criteria were as follows: (1) pathologically confirmed NB; (2) ≤ 18 years at first diagnosis; (3) A 18F-FDG PET/TC scan was made before operation or biopsy (within 4 weeks) (4) without any radiotherapy, chemotherapy or surgical treatment received before the 18F-FDG PET/CT examination included or of interest; (5) complete clinical information (Laboratory examination and bone marrow biopsy results); (6) available MNA data. Subsequently, 50 patients were excluded, including 27 patients without complete clinical information, and 23 patients had the above treatment at first diagnosis. According to the result of biopsy or surgery, there were 65 patients with MNA and 39 patients without MYCN amplification (Wild). All patients were assigned to the training and validation cohorts at random in a 7:3 ratio. Figure 1 depicts the flow chart for patient selection.

The flow chart for patient selection

The baseline data of each patient were obtained by reviewing the medical records and 18F-FDGPET/CT imaging. Clinical factors included age, gender, neuron-specific enolase (NSE), serum ferritin (SF), lactate dehydrogenase (LDH), vanillylmandelic acid (VMA) and homovanillic acid (HVA) level in a 24-h urine sample.

The radiological features (Table 1) of all patients were evaluated on a workstation (syngo.via, Siemens) by two nuclear medicine physicians with 5 and 10 years of experience in paediatric oncology diagnosis, respectively. But they were blinded to the clinical and histopathological diagnosis. In the event of a disagreement, a consensus was reached after further discussion. 18F-FDGPET/CT radiological features included International Neuroblastoma Risk Group Staging System (INRGSS), anatomical compartment, infiltration across the midline, calcification, and necrosis. Meanwhile, three conventional PET parameters were extracted from primary tumours (maximum standard uptake values (SUVmax), metabolic tumour volume (MTV), and total lesion glycolysis (TLG). Table 1 summarises the clinical factors and radiological features of the patients in the training and validation cohorts.

MNA was determined using FISH from paraffin-embedded tissue obtained by biopsy or surgery at the time of initial diagnosis, as previously published [12]. MNA was defined as a > fourfold increase in signals, according to the European Neuroblastoma Quality Assessment group [13, 14].

All patients were performed with a full body (from apex to toe) 18F-FDGPET/CT scanner (Biograph mCT-64 PET/CT; Siemens) [15]. They were asked to fast for at least 6 h and cut back on high-intensity exercise for at least 24 h before the injection. 18F-FDG (provided by Beijing Atomic High-tech Company) were injected intravenously 40–60 min before PET/CT scan. First, anatomical reference and attenuation correction were performed using low-dose CT scans, corrected with 120 kV tube voltage and auto-modulated tube current. CT image parameters are as follows: resolution 0.586 mm × 0.586 mm, slice thickness 2 mm, matrix size 512 × 512. PET scan was performed immediately after whole-body CT scan for 2 min per bed. The ordered subset expectation maximisation (OSEM) time-of-flight (TOF) algorithm was used to reconstruct PET images. PET image parameters are as follows: resolution 4.07 mm × 4.07 mm, slicer thickness 3 mm, matrix size 200 × 200.

Tumour segmentation and feature extraction were done as follows: to ensure the quality of the extracted radiomics features, 3D Slicer (version 4.10.2, funded by the National Institutes of Health) was used for medical image registration. The primary tumour delineation was performed using fixed SUV threshold method. According to the result of previous studies, 40% of SUVmax is set as the threshold for the images [16,17,18]. In this method, 3-D contours are drawn around voxels equal to or greater than 40% SUVmax. For the volume of interest (VOI) containing more than one cluster, the cluster which has maximum uptake intensity and volume is selected. A manual verification after automatic segmentation was performed; special attention was paid to tumour located near the urinary bladder due to intense physiological urinary tracer activity. If the VOI was found to be incorrect, additional manual adjustments were required. The VOIs included the lesion’s calcification and necrosis areas [19]. To minimise between-observer differences [20], each VOI was confirmed by two children’s nuclear medicine doctors (Q.L.D. and W.W.). For each precisely segmented VOI, the radiomics signature in the VOI was automatically extracted using radiomics in the open-source Python package (https://pyradiomics.readthedocs.io/en/Latest/). In each VOI, 1720 radiomics signatures were extracted from PET and CT images. These signatures include the following: (1) first-order features, (2) shape features, (3) and texture signatures (including grey-level co-occurrence matrix signatures (GLCM); grey-level dependence matrix signatures (GLDM); grey-level size zone matrix signatures (GLSZM); grey-level run-length matrix signatures (GLRLM); and neighbouring grey-tone difference matrix signatures (NGTDM)); we used Laplacian of Gaussian (LoG, sigma= 1, 3, 5, 7) and wavelet filtering to extract texture features.

Three months later, 40 patients were randomly selected from the training cohort to evaluate the reproducibility and robustness of the signature extraction process, and the nuclear medicine doctors (Q.L.D.) divided the data again and constructed a re-divided cohort. A value greater than 0.80 indicates good agreement when calculating intra/interclass correlation coefficients (ICCs).

Signature selection was done as follows: (1) use the z-score method to standardise all radiomics signatures in the training cohort; (2) Mann-Whitney U test retain signatures with p values less than 0.05; (3) Spearman correlation analysis and remove signatures with a correlation coefficient less than 0.9; (4) to find the most relevant predictive signatures, the least absolute shrinkage selection operator (LASSO) was used. Radiomics features underwent a multi-step selection process to overcome the limitations of traditional logistic regression methods, namely overfitting and multicollinearity problems in modelling high-dimensional radiomics signatures. The workflow was presented in Fig. 2.

Radiomics signature workflow

Constructing radiomics signatures was done as follows: after removing the redundant signatures, we feed the last set of radiomics signatures into the classifier to create a radiomics feature, which was used for biological assessment. In this study, we evaluated three classifiers: logistic regression (LR), decision tree (DT) and support vector machine (SVM). All classifiers choose the best performing model by using 10-fold cross validation in training cohort. To evaluate the performance of different radiomics models, the area under the curve (AUC), sensitivity, specificity, and accuracy were calculated using the receiver operating characteristic (ROC) curve, and the best model was the selected radiomics model with the highest AUC.

The clinical and radiological features were chosen primarily for their association with the MNA [21]. First, univariate analysis was used to identify clinical and radiological features that differed significantly from the MNA in the training cohort; then, multiple logistic regression analysis was used to identify the most relevant variables. Following multivariate analysis, clinical and radiological features independently related to MNA were used to develop a clinical-radiology (C-R) model. The radiomics model with the highest AUC outputs probability values for everyone, which are combined with clinical-radiological parameters to construct a multivariable logistic regression model (clinical-radiological-radiomics; C-R-R) and calculate the diagnostic efficiency of the model. The calibration curve and the Hosmer-Lemeshow test [22] were used to assess the model’s goodness of fit. To evaluate the clinical effectiveness of the model, a decision curve analysis (DCA) was used for the training and validation cohorts to calculate the net benefits under the threshold probability.

Statistical analyses were performed using R (version 4.1.0, Statistical Calculation Basics). A two-sided p value of less than 0.05 was considered statistically significant. Differences in all clinical features between two groups were assessed by independent samples t test, Mann-Whitney U test, and chi-squared or Fisher’s exact tests, as appropriate. DeLong test was used to compare the differences in AUC values between models. Accuracy, specificity, and sensitivity were calculated based on the cut-off value of the maximum Youden index.