Deep learning radiomic analysis in the prediction of MYCN status and survival outcome in children with neuroblastoma

Abstract

BACKGROUND

Neuroblastoma is the most common extracranial solid tumor in childhood, and its prognosis is strongly influenced by MYCN amplification. Non-invasive imaging biomarkers using deep learning (DL)-based radiomic analysis could capture whole-tumor characteristics to improve preoperative risk stratification and treatment planning.

AIM

To evaluate the performance of a DL-based strategy on contrast-enhanced computed tomography to predict the presence of MYCN amplification and clinical outcomes, event-free survival in patients with neuroblastoma.

METHODS

All 103 eligible patients were included retrospectively between 2008 and 2015, and assigned into the training cohort (n = 72) from one institution and the testing cohort (n = 31) from the other institution. We extracted DL-based features on pretrained convolutional neural networks via transfer learning automatically, which were classified by a support vector machine. A DL-based signature was formed from the DL-based model with the optimal area under the receiver operating characteristic curve (AUC) in the testing cohort. An integrated nomogram model, including significant clinical variables and the DL-based signature, was constructed for the prediction of histological patterns. Survival analysis of MYCN amplification identified by a nomogram predicted and histopathological results were evaluated in the prediction of tumor-associated events.

RESULTS

The nomogram model for MYCN amplification represented great prediction and discriminative performance with an AUC of 0.959 and accuracy of 96.7% in the training cohort and an AUC of 0.819 and accuracy of 74.7% in the testing cohort. Predicted probabilities ≤ -1.996 was considered as non-amplified MYCN status, and predicted probabilities > -1.996 was considered as amplified MYCN status. The non-amplified and amplified MYCN status showed a significant survival difference in event-free survival, categorized by both nomogram predicted and histopathological results.

CONCLUSION

The integrated clinical-DL model achieved accurate abilities in the identification of MYCN amplification before surgery, which might optimize therapeutic strategy and improve patients’ survival in neuroblastomas.

Key Words: Neuroblastoma; MYCN amplification; Deep learning; Radiomics; Computed tomography; Event-free survival

Core Tip: We constructed a deep learning (DL)-based radiomics signature on computed tomography, which had the ability to identify MYCN amplification in neuroblastoma. Integrating the DL-based radiomics signature and clinical predictors, the nomogram model showed improvement in the prediction of MYCN amplification in neuroblastomas. The DL-based radiomics signature was found to be associated with disease-specific events of neuroblastomas significantly after radical resection.

INTRODUCTION

Neuroblastoma represents the highest incidence among extracranial solid tumors in childhood, deriving from the developing sympathetic nervous system[1]. The prognosis of neuroblastoma depends on various factors, such as age at diagnosis, tumor stage, and histopathological characteristics, especially MYCN amplification[2]. The presence of MYCN, in about 20% of neuroblastomas, implies high aggressiveness and poor prognosis with the necessity of systematic treatments on the basis of surgical resection[2,3]. Nowadays, the identification of MYCN amplification is based on histopathological specimens from open or needle biopsies[4], which represents a weakness in inevitable errors from partial sampling on heterogeneous tumors[5] and invasive injuries from sampling procedures[6]. Therefore, an accurate detection of MYCN amplification is warranted with non-invasive approaches for improving therapeutic regimens.

Imaging materials have significant potential in representativeness of tumor characterization, and computed tomography (CT) and magnetic resonance imaging are recommended for diagnosis and staging of neuroblastomas in routine examinations. The contrast-enhanced CT examinations can be performed by fast acquisitions without sedation in children. The CT images of the whole tumor volume are feasible for the identification of MYCN amplification with macroscopic evidence of intra-tumoral heterogeneity in neuroblastomas. The emerging radiomics techniques can interpret the underlying characteristics of the whole tumor with high-throughput quantitative features to associate with multiple clinical outcomes[7]. With integration with advanced deep learning (DL) methods, DL-based radiomics recognizes accurate and consistent imaging features on images directly via convolutional neural networks (CNNs) to indicate associations with clinical endpoints[8]. The approach, including feature extraction by DL-CNNs and clinical outcomes prediction via supervised machine learning methods, achieves great performance in both predictive accuracy and computational costs for clinical tasks[9]. The prediction tool on DL-based radiomics analysis has the potential to identify MYCN amplification before invasive interventions in order to instruct the subsequent treatment and improve survival outcomes. In the present study, the objective was to evaluate the ability of a DL-based imaging-derived tool in the identification of MYCN amplification and in the prediction of survival outcomes on contrast-enhanced CT images in patients with neuroblastoma.

MATERIALS AND METHODS

Patients

We recruited 103 patients retrospectively diagnosed with neuroblastoma who underwent radical resection between January 2008 and December 2015 in two institutions. The inclusion criteria were as follows: (1) Age less than 18 years old at diagnosis; (2) Confirmation of MYCN amplification detection by histopathological specimens; (3) Receiving radical resection; and (4) Presence of contrast-enhanced CT images before interventions. The inclusion patients from two institutions were assigned to the training cohort (72 patients) and testing cohort (31 patients), respectively. The institutional review boards of two institutions approved this retrospective study and waived the requirement to obtain written informed consent from all participants. Demographic and clinical data were collected for each patient and comprised age at diagnosis, gender, tumor stage, and histopathological results. The tumor stage was defined according to the International Neuroblastoma Risk Group Staging System (INRGSS)[10]. The differentiation results were divided into two groups: Poorly differentiated or undifferentiated neuroblastoma and differentiated neuroblastoma[11]. The CT imaging protocol parameters in neuroblastomas were listed in the Supplementary material.

Clinical outcomes and follow-up

The primary outcome was to identify whether there was the presence of MYCN amplification in histopathological specimens. MYCN status was measured by fluorescence in situ hybridization technique, using a dual-color probe and a chromosome 2 centromere enumeration probe as a reference (Vysis LSI N-MYC SpectrumGreen/CEP 2 SpectrumOrange Probes, Abbott Laboratories, IL, United States). The presence of MYCN amplification was defined as a fourfold increase in the MYCN signal number compared with the reference probe[4]. Event-free survival (EFS), the secondary outcome, was defined as the time from the date of operation to the date of any events, such as recurrence, metastasis, death, or the last follow-up. The follow-up schedule was performed every 3 months for the first year, then every 6 months from the second year to the fifth year, and annually thereafter. After surgery, each follow-up consisted of CT scanning of primary lesions, chest X-ray, and blood tests. The last follow-up was censored in December, 2020.

Tumor delineation and preprocessing

The regions of interest (ROIs) of the primary tumor were delineated by two experienced radiologists in contrast-enhanced CT images separately using ITK-SNAP software[12]. For patients absent of MYCN amplification with more than one primary lesion in CT images, the largest lesion was depicted with the consensus of two radiologists. For patients present of MYCN amplification with more than one primary lesion in CT images, the lesion found with MYCN amplification was selected for ROI delineation. The current study used expert manual ROI delineation because pre-existing, multi-institutional test-retest data were not available. With the masked delineation of ROIs for primary tumor, we adjust the size of consecutive axial slices of primary lesions into 224 mm × 224 mm for the input layer of the pre-trained CNN models using a bounding box covering the whole tumor area. The resized images with one axial slice of the maximum area as image channels would be candidate of the CNNs for feature extraction of 2-dimensional (2D) DL features. And the resized images of the whole tumor volume as image channels would be candidates for the pre-trained DL algorithms for feature extraction of 3-dimensional (3D) DL features.

DL-based features extraction

Pretrained CNN architectures were applied to contrast-enhanced CT images for the extraction of 2D and 3D DL-based features, respectively. There were six pre-trained CNN algorithms used in this study, including Xception[13], VGG16[14], VGG19[14], ResNet50[15], InceptionV3[16], and InceptionResNetV2[17]. These pre-trained CNNs were commonly trained and applied to the large-scale and well-annotated ImageNet database[18]. These extracted DL-based features from different pre-trained CNNs would be used for the construction of DL-machine learning approaches to identify MYCN amplification at the next step. A detailed description of DL features extraction was shown in the Supplementary material. For ResNet50, we extracted features from several stages, including early, mid, and the last convolutional block, Res5c, and constructed DL-support vector machine (SVM) models for each layer’s features to empirically assess which layer provided the best discriminative information. Since the mechanism for the prognostication of DL-based features was not clear, considering the complexity of pre-trained CNNs, we interpreted ROIs recognized by DL-CNNs into heatmaps. Guided Gradient-weighted Class Activation Mapping visualized the output of DL-CNNs in the last convolutional layer[19], which highlighted specific subregions CNNs are concerned with for the generation of DL-based features.

MYCN amplification-related DL-based signatures

We constructed DL-machine learning models for discrimination between the absence and presence of MYCN amplification based on extracted 2D and 3D numeric DL-based features. Three kinds of DL-machine learning models were developed on 2D, 3D, and combined DL-based features, respectively. We selected the top 20% best DL-based features in the prediction of MYCN amplification calculated by univariate analysis in the training cohort. Next, we implemented a wrapper feature selection method regarding the recursive feature addition algorithm to select the predictive features with a high area under the receiver operating characteristic (ROC) curve (AUC) value. SVM with a radial basis kernel function as a classifier was used to construct DL-SVM models on selected DL-based features[20]. All DL-SVM models were evaluated by AUC in the training and testing cohorts. We chose the best DL-SVM model with the highest AUC of the testing cohort among all DL-SVM models in association with MYCN amplification to generate the DL-based signature.

Development of MYCN amplification-related clinical-DL model

All clinical variables were evaluated by univariate and multivariate logistic regression in the training cohort. The univariate analysis selected variables with a significant P-value to enter the multivariate regression model. In the multivariable logistic regression, these clinical variables with a significant P-value were considered as independent clinical predictors for the identification of MYCN amplification. The clinical-DL model was constructed by significant clinical predictors and the MYCN amplification-related DL signatures. The clinical-DL model was formulated based on the independent predictors of multivariate regression. This model is based on proportionally converting each regression coefficient in multivariate logistic regression to a 0-point to 100-point scale. The effect of the variable with the highest b coefficient (absolute value) is assigned 100 points. The points are added across independent variables to derive total points, which are converted to predicted probabilities (Pi). Pi indicated the predicted probabilities of the clinical-DL model in association with MYCN amplification.

MYCN amplification-predicted EFS

All clinical variables for EFS in the training cohort were evaluated by Cox proportional hazard analysis, where the univariate analysis selected the variables with a significant P-value, entered the multivariate regression model, and the multivariate Cox regression identified the prognostic factors with significant statistical results. The Pi of all patients was categorized into the low-score and high-score groups according to the optimal cut-off points from the training cohort on EFS, respectively, via the X-tile software (version 3.6.1; Yale University School of Medicine, New Haven, CT, United States)[21]. The category with low-score Pi was considered as the absence of MYCN amplification group, and the category with high-score Pi was considered as the presence of MYCN amplification group. Survival analysis was performed between the absent and present MYCN amplification groups by Pi-predicted and histopathological results, respectively.

Statistical analysis

As for the distribution of clinicopathological parameters, we applied Fisher’s exact test or χ² test for categorical data and non-parametric Mann-Whitney test for continuous data to identify the difference between the training and testing cohorts. For the evaluation of predictive performance of DL-SVM models, we applied the ROC curves and their value of AUC, accuracy, sensitivity, and specificity in the training and testing cohorts. The calibration performance was measured by the Hosmer-Lemeshow test and visualized by calibration plots[22]. Discrimination ability was quantified by Harrell’s concordance indices (C-index)[23]. The nomogram of the MYCN amplification-related clinical-DL model was applied, evaluated by ROC curves, and calibration plots. Survival curves with the Kaplan-Meier (KM) method were generated and compared by a 2-sided log-rank test. Variables that reached statistical significance in the univariate analysis were considered for the multivariate model. A two-tailed P-value less than 0.05 was considered statistically significant. The whole statistical analysis and graphic production were completed by Python (version 3.8) and R (version 3.6.1). All involved packages in this study are listed in the Supplementary material.

RESULTS

Demographic characteristics

The whole scheme was depicted, including feature extraction, MYCN amplification prediction, and EFS prediction (Figure 1). A total of 103 patients were included in this study that 72 patients from one institution were assigned to the training cohort, and 31 patients from the other institution were assigned to the testing cohort. There were no significant differences found in baseline variables and clinical outcomes between the training and testing groups (P > 0.05). Demographic characteristics are shown in Table 1 in detail. There were 11.7% (12/103) neuroblastoma patients found presence of MYCN amplification on histopathological specimens.

Figure 1 A general flowchart of data analysis. A: Deep learning (DL)-based feature extraction was performed on contrast-enhanced computed tomography images of neuroblastoma by pre-trained convolutional neural network algorithms, which transformed the original images into numeric DL-based features; B: The MYCN amplification prediction was approached by DL-based features via DL-based signature generation and integrated nomogram model construction. The DL-based signature was generated on the DL-based model with the largest area under the receiver operating characteristic curve value in the testing cohort. The integrated nomogram model combined significant clinical variables and the DL-based signature; C: The survival analysis in the prediction of event-free survival was evaluated on MYCN amplification identified by nomogram-predicted and histopathological results. Predicted probabilities (Pi) are the total points that derive from independent variables in the nomogram model for differentiating non-amplified and amplified MYCN status. The optimal cut-off point of Pi was -1.996, defined by X-tile software, which indicated Pi ≤ -1.996 is considered as non-amplified MYCN status, and Pi > -1.996 is considered as amplified MYCN status. DL: Deep learning; SVM: Support vector machine; AUC: Area under the receiver operating characteristic curve; EFS: Event-free survival; Pi: Predicted probabilities.

Table 1 The demographic and clinical characteristics between training and testing cohorts, n (%).

Characteristic	All subjects (n = 103)	Training cohort (n = 72)	Testing cohort (n = 31)	P value
Age, median (IQR), months	24 (17)	24 (17)	25 (18)	0.701
Gender				0.905
Male	81 (78.6)	57 (79.2)	24 (77.4)
Female	22 (21.4)	15 (20.8)	7 (22.6)
MYCN amplification				0.518
Non-amplified	91 (88.3)	64 (88.9)	27 (87.1)
Amplified	12 (11.7)	8 (11.1)	4 (12.9)
INRGSS stage				0.706
L1	47 (45.6)	33 (45.8)	14 (45.2)
L2	56 (54.4)	39 (54.2)	17 (54.8)
Histological differentiation				0.288
Differentiated	38 (36.9)	26 (36.1)	12 (38.7)
Undifferentiated or poorly differentiated	65 (63.1)	46 (63.9)	19 (61.3)
Infiltrating across the midline				0.880
Absence	54 (52.4)	38 (52.8)	16 (51.6)
Presence	49 (47.6)	34 (47.2)	15 (48.4)
Calcification				0.960
Absence	32 (31.1)	22 (30.6)	10 (32.3)
Presence	71 (68.9)	50 (69.4)	21 (67.7)
DL-based signature, median (IQR)	-0.243 (0.154)	-0.244 (0.153)	-0.243 (0.229)	0.311
EFS, median (IQR), month	72 (78)	71 (78)	72 (78)	0.437

IQR: Interquartile range; INRGSS: International Neuroblastoma Risk Grouping Staging System; DL: Deep learning; EFS: Event-free survival.

MYCN amplification prediction

Three significant clinical variables were selected as independent predictors for discrimination between amplified and non-amplified MYCN status in the training cohort, including histological differentiation, presence of infiltration across midline, and calcification (Supplementary Table 1). Their AUC and corresponding 95% confidence interval (CI) were calculated with 0.433 (95%CI: 0.337-0.529) for histological differentiation, 0.581 (95%CI: 0.485-0.678) for presence of infiltrating across midline, and 0.505 (95%CI: 0.408-0.603) for presence of calcification in the testing cohort (Table 2).

Table 2 Diagnostic performance of clinical predictors, deep learning-based signature, and integrated nomogram model for prediction of MYCN amplification.

Model	Training cohort (n = 72)				Testing cohort (n = 31)
	AUC (95%CI)	Accuracy	Sensitivity	Specificity	AUC (95%CI)	Accuracy	Sensitivity	Specificity
Histological differentiation	0.560 (0.495-0.625)	47.2 (34/72)	62.5 (5/8)	45.3 (29/64)	0.433 (0.337-0.529)	48.4 (15/31)	25.0 (1/4)	51.9 (14/27)
Infiltrating across midline	0.603 (0.536-0.669)	65.3 (47/72)	62.5 (5/8)	65.6 (42/64)	0.581 (0.485-0.678)	61.3 (19/31)	50.0 (2/4)	63.0 (17/27)
Calcification	0.518 (0.450-0.587)	88.9 (64/72)	1.3 (1/8)	98.4 (63/64)	0.505 (0.408-0.603)	87.1 (27/31)	0.0 (0/4)	100.0 (27/27)
DL-based signature	0.958 (0.929-0.987)	91.7 (66/72)	87.5 (7/8)	92.2 (59/64)	0.803 (0.723-0.883)	77.4 (24/31)	75.0 (3/4)	77.8 (21/27)
Nomogram model1	0.959 (0.930-0.988)	95.8 (69/72)	87.5 (7/8)	96.9 (62/64)	0.819 (0.740-0.898)	74.2 (23/31)	75.0 (3/4)	74.1 (20/27)

¹The nomogram model integrated significant clinical predictors and the deep learning-based signature to predict MCYN amplification.

AUC: Area under the receiver operating characteristic curve; DL: Deep learning; CI: Confidence interval.

As for the performance of DL-SVM models for differentiation between amplified and non-amplified MYCN, the AUCs ranged from 0.873 to 0.966 for the training group and from 0.553 to 0.819 for the testing group (Supplementary Table 2). A total of 18 DL-SVM models, including 6 2D models, 6 3D models, and 6 combined models, were constructed based on selected features from different pre-trained DL-CNNs (Supplementary Table 3). The combined model with ResNet50 algorithm achieved the best discriminative performance considering the AUC value of the testing cohort among all DL-SVM models, with an AUC of 0.959, a C-index of 0.959, accuracy of 96.53%, sensitivity of 90.12%, and specificity of 97.34% in the training cohort, and an AUC of 0.819 (95%CI: 0.740-0.898), a C-index of 0.819, accuracy of 74.68%, sensitivity of 74.36%, and specificity of 74.72% in the testing cohort (Supplementary Table 2). For the ResNet50 CNN model, DL-based features extracted from different layers might exist differences since its complex DL-CNN architecture. We constructed DL-SVM models on DL features from earlier layers of the ResNet50 algorithm to explore the superiority of DL features from the last layer before the fully connected layer, Res5c, in the classification of amplified and non-amplified MYCN (Supplementary Table 4). The model constructed using features from the Res5c layer outperformed models built from earlier layers, as shown in Supplementary Table 4, where the DL-SVM using Res5c features achieved higher AUC and C-index in both training and testing cohorts than DL-SVMs built from earlier blocks, suggesting Res5c features capture more task-specific, high-level image patterns relevant to MYCN amplification. The results demonstrated the advantages of the current extraction strategy from the layer of Res5c in the ResNet50 algorithm. ResNet50 architecture was applied to generate feature maps highlighting important subregions for the generation of the output (Figure 2). As shown in the feature heatmaps, the valuable areas were different in different MYCN status for the extraction of specific feature patterns, which were interpreted for further clinical view.

Figure 2 Feature heatmaps of representative histological patterns in bladder cancer on the deep learning ResNet50 algorithm via the Guided Grad-Class activation mapping. A and B: The original computed tomography images and their corresponding feature heatmaps were shown in left and right sides, respectively. The red color highlighted the region of interest to classify MYCN amplification. Different subregions were recognized and highlighted in different MYCN statuses, including amplified MYCN status (A) and non-amplified MYCN status (B).

The DL-based signature was generated based on the combined ResNet50 DL-SVM model. Integrating independent clinical predictors and the DL-based signature (Figure 3A), an integrated nomogram model for prediction of histological patterns was constructed, including significant clinical variables and the DL-based signature (Figure 3A) with an AUC of 0.959 and accuracy of 96.7% in the training cohort, and an AUC of 0.819, accuracy of 74.7%, sensitivity of 74.4%, and specificity of 74.7% in the testing cohort (Table 2). This nomogram model represented great classification performance by ROC curves (Figure 3B) and consistent calibration capacities (Figure 3C) in both training and testing cohorts. The Pi derived from the nomogram model was calculated with the formula: Pi = 0.823 × Histological differentiation + 0.280 × presence of infiltrating across midline + 0.368 × presence of calcification + 5.684 × DL-based signature - 2.577.

Figure 3 Evaluation of predictive performances for the integrated nomogram model for the prediction of MYCN amplification in neuroblastomas. A: Nomogram model combining significant clinical variables and the deep learning (DL)-based signature. The DL-based signature was generated from the best DL-based model considering the area under the receiver operating characteristic curve (AUC) of the testing cohort among all models; B: Receiver operating characteristic curves for the predictive performance of the nomogram model in the training and testing cohorts, respectively. The AUC of the training cohort was 0.959, and the AUC of the testing cohort was 0.819; C: Curves of the calibration analysis for the nomogram model in the training and testing cohorts, respectively. AUC: Area under the receiver operating characteristic curve; ROC: Receiver operating characteristic; DL: Deep learning.

EFS prediction

According to the results of multivariate Cox regression, age at diagnosis, INRGSS stage, histological differentiation, presence of infiltrating across midline, presence of MYCN amplification, and nomogram model-predicted probability were considered as independent predictors for EFS (Supplementary Table 5). The hazard ratio and corresponding 95%CI were calculated with 1.011 (95%CI: 1.002-1.021) for age at diagnosis, 2.674 (95%CI: 2.117-3.377) for INRGSS stage, 1.319 (95%CI: 1.046-1.665) for histological differentiation, and 1.541 (95%CI: 1.210-1.963) for presence of Infiltrating across midline (Supplementary Table 5). The nomogram model-predicted probability was stratified with the optimal cut-off points of -1.996 for recurrence-free survival via X-tile software from the training cohort. Pi ≤ -1.996 was considered as non-amplified MYCN status, and Pi > -1.996 was considered as amplified MYCN status. The KM survival analysis for the histopathological non-amplified and amplified MYCN groups showed the mean survival of 83.476 months (95%CI: 79.401-87.550 months) vs 63.938 months (52.568-75.308 months) in the training cohort, and the mean survival of 85.458 months (95%CI: 79.308-91.609 months) vs 82.538 months (66.945-98.130 months) in the testing cohort, with the P-value of 0.031 in the log-rank test (Table 3). The KM survival analysis for the model-predicted non-amplified and amplified MYCN groups showed the mean survival of 85.504 months (95%CI: 81.349-89.659 months) vs 60.239 months (50.833-69.644 months) in the training cohort, and the mean survival of 83.366 months (95%CI: 76.510-90.222 months) vs 89.230 months (78.847-99.614 months) in the testing cohort, with the P-value of 0.002 in the log-rank test (Figure 4 and Table 3).

Figure 4 Kaplan-Meier survival curves on event-free survival of MYCN amplification status for patients with neuroblastoma in the training and testing cohorts. A: Kaplan-Meier survival curves of nomogram model-predicted results, non-amplified vs amplified MYCN status, in the training and testing cohorts, respectively. Predicted probabilities (Pi) are the total points that derive from independent variables in the integrated nomogram model for differentiating non-amplified and amplified MYCN status. The optimal cut-off point of Pi was -1.996, defined by X-tile software, which indicated Pi ≤ -1.996 is considered as non-amplified MYCN status, and Pi > -1.996 is considered as amplified MYCN status; B: Kaplan-Meier survival curves of histopathological results, non-amplified vs amplified MYCN status, in the training and testing cohorts, respectively.

Table 3 Kaplan-Meier analysis of histologically identified and nomogram model-predicted MYCN amplification for the prediction of long-term survival in neuroblastomas.

Variable	Subgroup	Training database		Testing database		Log rank test
		Mean survival time	95%CI	Mean survival time	95%CI
Histological MYCN amplification1	Non-amplified	83.476	79.401-87.550	85.458	79.308-91.609	0.031
	Amplified	63.938	52.568-75.308	82.538	66.945-98.130
Predicted MYCN amplification2	Non-amplified	85.504	81.349-89.659	83.366	76.510-90.222	0.002
	Amplified	60.239	50.833-69.644	89.230	78.847-99.614

¹The histological MYCN amplification was identified by postoperative histopathological specimens.

²The predicted MYCN amplification was derived from two categories from the nomogram model-predicted probabilities based on the optimal Youden index of receiver-operating characteristic analysis, where the lower group indicated non-amplified MYCN status and the higher group indicated amplified MYCN status.

CI: Confidence interval.

DISCUSSION

In this study, we aimed to develop predictive models for MYCN amplification and survival outcomes in patients with neuroblastomas. We converted CT images into quantitative signatures via DL-based methods and constructed an integrated nomogram model with great preoperative discriminative capacities for MYCN amplification. Additionally, the nomogram model predictions were independently associated with disease-specific events of neuroblastomas after radical resection. Our results suggested that the DL-based models for the prediction of MYCN amplification on contrast-enhanced CT might instruct individualized treatment and survival surveillance of neuroblastomas.

The correct identification of MYCN amplification before surgery is pivotal in order to establish an appropriate therapeutic strategy and to improve survival outcomes of patients with neuroblastomas. Indeed, accurate preoperative diagnosis of patients with amplified MYCN status needs multi-model approaches on clinicopathological information, which guides individualized therapeutic regimens. In the present study, the DL-based radiomics signature on contrast-enhanced CT images provided useful information for detecting MYCN amplification. Nowadays, the acquisition of histopathological specimens from needle core biopsy is the mainstay for the determination of MYCN status to assist clinicians in establishing subsequent treatment[24]. The specimens from the open or needle biopsies lack representativeness of the entire bladder cancer lesion due to the small amount of the specimen and sampling bias[5]. Previous literature showed different MYCN amplification status between needle-biopsy and resection specimens[25]. These studies reported that MYCN amplification status defined on needle core-biopsy samples was slightly concordant with evidence determined by resection specimens of the whole tumor due to the small specimen size hard to capture important and specific morphological features such as high vascularity induced by MYCN amplification[26]. The needle-biopsy samples represented little feasibility for the identification of MYCN amplification, whereas the surgical specimens might result in too much invasive damage for children[27]. Because of the inaccurate reflection of tumor characterization on insufficient samples, we need other tools to evaluate extensive regions for evidence about histological inclination on neuroblastomas, as the preoperative CT images have the ability to provide useful data on the whole tumor lesions.

The correct identification of MYCN amplification could improve the patients’ survival and quality of life. In cases of amplified MYCN neuroblastomas, patients show strong tendencies for rapid progression in local infiltration, metastasis, and relapse[8]. If any preoperative histological evidence indicated the presence of MYCN amplification, the subsequent treatment would be inclined to neoadjuvant chemotherapy rather than surgical resection for better control of local lesions[28]. The computer-aided diagnosis (CAD) on imaging data has the potential to provide information about MYCN amplification based on CT images. The application of DL-based techniques was expected to achieve great prediction performance with well concordance with findings on surgical resection of the whole tumor. In our main results, the DL-based signature in the present study achieved great discrimination abilities in the classification of non-amplified and amplified MYCN status with an AUC of 0.958 (95%CI: 0.929-0.987) in the training cohort and 0.803 (95%CI: 0.723-0.883) in the testing cohort. The integration of the DL-based signature and prognostic clinical variables had a slight impact on discriminative abilities with an AUC of 0.959 (95%CI: 0.930-0.988) in the training cohort and 0.819 (95%CI: 0.740-0.898) in the testing cohort. The DL-based imaging analysis indeed yielded strong classification capacities to differentiate patients with MYCN amplification from patient absent of MYCN amplification on preoperative CT images, which might assist in establishing individualized treatment before operation.

In the present study, we found that the presence of MYCN amplification was significantly associated with poor EFS in both nomogram-predicted and histopathological results. According to the previous studies, the impact of MYCN status was not consistent on survival outcomes of patients with neuroblastoma. The findings in early-stage neuroblastomas showed agreement with our study that VHs were significantly associated with poor EFS[29]. On the contrary, some research indicated that there existed little robust association between the presence of MYCN amplification and survival outcomes in neuroblastomas in advanced neuroblastomas[30]. The effects of MYCN amplification might have a strong association with treatments based on tumor stage. These conflicting findings underline the importance of exploring the actual impact of MYCN amplification on survival outcomes in patients with neuroblastoma. These controversial conclusions might attribute the correct detection rates of MYCN amplification in pathological specimens, affected by sampling size and acquisition methods. The proper selection of patients with amplified MYCN status who are most likely to benefit from systemic therapy is pivotal to avoid over- and under-treatment.

In this study, we expected to explore further evidence to classify MYCN amplification by our non-invasive CAD systems. The diagnostic performance of our CAD systems was reliable and cost-effective, which might be an effective complementary measure before surgery. The practicability of our DL-based CAD systems was demonstrated in solving one specific medical task with respect to the satisfactory predictive results. The actual procedure of these CAD systems was hard to interpret directly for understanding. Class activation mapping heatmaps were used to translate the important subregions recognized by CAD systems. According to Figure 2, a different area was recognized for differentiation between amplified and non-amplified MYCN status. These differences demonstrated the potential of CAD systems in the detection of spatial heterogeneity of tumoral and peritumoral microenvironment.

Some limitations existed in the present study. First and foremost, the retrospective design of the present study was companied with the presence of inherent selection bias and the absence of standardization of therapeutic protocols. Since the small testing sample, an external validation plan based on a multi-center prospective study is being conducted with pre-specified endpoints, sample size calculation, model update, and performance stability measures. Second, the ROI contour-based method might be insufficient in assurance of DL-based feature robustness and repeatability compared with the test-retest method[31]. Since no test-retest scans were available in this retrospective cohort, the robustness of DL-derived features could be affected by contour variability and acquisition differences. Our research team is trying to apply the test-retest method in a prospective study as the next step. In our planned prospective study, we will obtain dedicated test-retest CT scans on a subset of patients and use phantom experiments to quantify repeatability and derive robust DL-features using the intraclass correlation coefficient and concordance metrics, and exclude unstable features prior to model development. Third, the heterogeneity between the original pre-trained CNN bases and target databases was hard to avoid, even though we adopted transfer learning to mitigate it. A DL-based model from scratch, constructed on a large database, is an ideal solution for excellent generalizability and clinical practicability of DL-based CAD methods. Last but not least, we considered the mixed cases as amplified MYCN status once the appearance of morphological MYCN amplification occurred without quantifying the amount and percentage. In this retrospective study, mixed or focal amplification was classified as positive due to limitations of available pathology reports. An “all-or-nothing” dichotomy might misrepresent intratumoral heterogeneity and could bias model predictions and survival associations. A further comprehensive evaluation of image-to-genetic associations of neuroblastomas will be conducted for prediction of survival outcomes. In future work, we will collect detailed pathology reports with quantitative measures of amplification, including percentage of amplified cells and copy-number estimates, perform sensitivity analyses excluding focal/mixed cases to assess robustness, and explore multi-class outcomes or continuous target labels like copy-number or amplification fraction to capture heterogeneity more accurately.

The DL-based nomogram has been envisioned as a preoperative decision-support and triage tool rather than an immediate replacement for histopathology. In practice the workflow would be: (1) All children undergo routine preoperative contrast-enhanced CT; (2) Images are processed through the trained DL pipeline and the nomogram produces a Pi for MYCN amplification; (3) Patients with high predicted probability would be discussed in the multidisciplinary team to consider expedited molecular testing, neoadjuvant therapy, or prioritized comprehensive biopsy sampling to confirm MYCN amplification; and (4) Patients with low predicted probability would follow the standard diagnostic pathway. To justify replacing biopsy for treatment decisions would require high-level evidence: Multi-center prospective diagnostic accuracy studies showing high sensitivity/specificity with narrow CIs, followed by impact studies demonstrating that decisions based on the model improve outcomes or reduce harm. Until such evidence is obtained, the model would be recommended to triage and augment decision-making, not to abolish tissue-based testing. Otherwise, for our planned prospective segmentation and standardization, we will implement a semi-automated segmentation pipeline using a convolutional segmentation network like U-Net and nnU-Net, trained on the current cohort’s expert contours. The prospective workflow will be automated segmentation, rapid review by two trained radiologists, and adjudication by a third senior reader if not achieved. Inter-reader reliability and test-retest repeatability will be assessed on an initial calibration set, and only features with high repeatability will be considered for final model building to ensure the robustness required for clinical deployment.

CONCLUSION

In conclusion, we demonstrated that the DL-based signature showed notable advantages in the identification of MYCN amplification over existing clinical variables for neuroblastomas. The integrated nomogram model achieved a great differentiation performance for the classification of non-amplified and amplified MYCN status. The integrated diagnostic tool represented a novel manner with high efficiency and without manual tumor delineation to assist the preoperative findings and enhance the correct diagnostic rate of MYCN amplification before surgery. MYCN amplification identification could facilitate predicting tumor-associated events and optimizing treatment for patients with neuroblastoma.