Advancing Neuroblastoma Diagnosis: A Comprehensive Review of Computational Technologies and Methodologies for Data Analysis and Pattern Extraction

NB is the most common extracranial solid tumour in paediatric patients, originating from embryonic neural crest cells known as neuroblasts. Typically, these immature cells differentiate into functional nerve cells. Still, in NB, they undergo uncontrolled proliferation, leading to tumour formation primarily in the adrenal glands, which regulate hormone production, blood pressure, and other vital functions [1]. However, NB can also develop in different regions of the sympathetic nervous system, including the abdomen, chest, and neck. In its early stages, NB may be asymptomatic, making early detection challenging. As the disease progresses, tumour metastasis occurs through hematogenous and lymphatic spread, affecting distant organs such as the bones, liver, lungs, and bone marrow, significantly complicating prognosis [1, 2]. The heterogeneous nature of NB leads to widely varying clinical outcomes, ranging from spontaneous regression in low-risk cases to aggressive, treatment-resistant tumours in high-risk patients.

NB predominantly affects infants and children under the age of five, accounting for 15% of paediatric cancer-related deaths. Prognosis is influenced by age at diagnosis, tumour stage, MYCN amplification status, and histopathological classification. Despite advancements in molecular profiling and targeted therapy, early and accurate diagnosis remains a significant challenge due to overlapping clinical features with other paediatric malignancies [3].

To address these challenges, computational approaches, including ML and deep learning DL, have emerged as powerful tools for NB detection, risk stratification, and treatment prediction. This review explores state-of-the-art computational methodologies integrating multi-omics, histopathological imaging, and clinical data to enhance early diagnosis, prognosis, and personalized treatment strategies.

1.1 Statistical study of Neuroblastoma

NB is the predominant tumour affecting the sympathetic nervous system, accounting for 97% of cases. It is also the most prevalent form of cancer among infants, with an average diagnosis age of 17 months [1]. It constitutes 15% of all paediatric cancer-related deaths [2]. The yearly occurrence of NB in the United States is approximately 650 cases, translating to an incidence rate of 10.2 cases per million children (or 65 cases per million infants). This rate has remained stable over time, with a negligible variation of 0.4% [3]. Despite a general decrease in the five-year mortality rate from 1975 to 2005, subgroup-specific mortality trends indicate significant variation in survival outcomes.

Globally, variations in detection rates exist due to differences in screening programs and healthcare infrastructure. For example, urinary catecholamine screening in Japan has led to earlier NB detection, potentially reducing mortality rates in high-risk cases. However, standardization challenges prevent the widespread adoption of such programs worldwide [4]. A graphical representation comparing incidence rates, survival trends, and regional variations can provide additional insight into how NB is managed across different populations.

1.2 Pathophysiological behaviour of Neuroblastoma

NB arises due to genetic and molecular alterations in neural crest cells, leading to aberrant cellular proliferation and differentiation failure. Several key biochemical and molecular markers have been identified in its pathophysiology:

● MYCN Amplification: In ~25% of cases, MYCN amplification is associated with rapid tumour progression and poor prognosis [4]. Patients with MYCN-amplified tumours often exhibit aggressive disease, early metastasis, and chemotherapy resistance.

● ALK Mutations: Anaplastic lymphoma kinase (ALK) mutations have been implicated in NB pathogenesis and represent a target for novel therapies.

● NTRK-1 & CD-44 Expression: While NTRK-1 expression is linked to low-risk tumours, CD-44 absence correlates with poor prognosis.

● DNA Index (Ploidy Status): Tumours with a high DNA index (hyperdiploidy) respond better to chemotherapy, whereas near-diploid tumours have a worse prognosis.

● Serum Biochemical Markers: Elevated levels of LDH, serum ferritin, and neuron-specific enolase (NSE) indicate a higher tumour burden and worse outcomes [5].

Approximately 90% of patients show increased urinary vanillylmandelic acid (VMA) and homovanillic acid (HVA) levels, aiding in biochemical screening and diagnosis. The heterogeneity of NB makes risk stratification essential for personalized treatment planning.

1.3 Developmental stages of Neuroblastoma

Assessing the tumour stage is a critical factor in determining the risk category and treatment strategy. Two primary staging systems are used.

1.3.1 International Neuroblastoma Staging System (INSS)

Introduced in 1986, the INSS relies on post-surgical tumour assessment (Figure 1). This staging method depends on the extent of surgical tumour resection, making it less applicable for children who are inoperable at diagnosis. 

● Stage 1: Localized tumour, completely resected, no lymph node involvement.

● Stage 2A: Localized tumour, incomplete resection, no lymph node involvement.

● Stage 2B: Localized tumour with regional lymph node involvement.

● Stage 3: Tumour crosses the midline or involves contralateral lymph nodes.

● Stage 4: Distant metastases present.

● Stage 4S: Metastases limited to liver, skin, and bone marrow in infants under one year, associated with better prognosis.

Figure 1 visually represents the INSS staging criteria, showing how tumour progression correlates with surgical removal feasibility and metastasis.

1.3.2 International Neuroblastoma Risk Group Staging System (INRGSS)

The INRGSS, established in 2005, is based on pre-treatment imaging (CT, MRI, and MIBG scans) rather than surgical outcomes. This approach enables risk stratification before surgical intervention.

● L1: Localized tumour, no Image-Defined Risk Factors (IDRFs).

● L2: Tumour with IDRFs but without distant spread.

● M: Distant metastases present.

● MS: Metastases limited to skin, liver, or bone marrow, often seen in infants under 18 months.

Figure 2 provides a detailed overview of the developmental stages of NB using the INRGSS framework. It also integrates MYCN amplification status and histopathological differentiation, which are key determinants of risk stratification in NB patients [1].

Figure 1. INSS staging criterion to access NB

Figure 2. Developmental stages of Neuroblastoma with risk associates

Table 1. Comparision on staging system on the basis of certain criterion

Table 1 provides a comparative summary of INSS vs. INRGSS, highlighting key differences in staging criteria, surgical requirements, and risk classification.

1.3.3 Shimada classification system for Neuroblastoma

The Shimada Classification System, developed by the International NB Pathology Committee, categorizes NB tumours based on histopathological differentiation and stromal composition. It is a key determinant of prognosis, particularly in distinguishing between favorable and unfavorable histology (UH).

● Favorable Histology (FH): Characterized by well-differentiated ganglion cells and abundant Schwannian stroma. More common in younger children and associated with a better prognosis.

● UH: Poorly differentiated or undifferentiated tumour cells with high mitosis-karyorrhexis index (MKI), typically observed in older children with aggressive disease.

Key criteria in the Shimada system include:

● Schwannian stroma content (stroma-rich vs. stroma-poor tumours).

● Differentiation level (differentiating, poorly differentiated, or undifferentiated NB).

● MKI (low, intermediate, or high).

● Age of the patient (critical in determining prognosis in conjunction with histology).

The Shimada classification is essential in risk stratification, guiding treatment decisions alongside INSS and INRGSS.

Figure 3 illustrates the Shimada classification standard, showcasing its role in predicting patient outcomes based on tumour histology [6].

1.4 Clinical strategies to deal with Neuroblastoma

The management of NB requires a multidisciplinary approach, integrating surgery, chemotherapy, radiation therapy, immunotherapy, and emerging AI-driven diagnostic and treatment strategies. Treatment choice depends on tumour staging, molecular characteristics, and patient-specific risk factors [7].

Surgical intervention is the first-line treatment for localized NB. In low-risk cases, complete surgical resection can be curative. However, for high-risk or advanced-stage tumours, surgery is often complemented by chemotherapy and radiation therapy. Chemotherapy is used in two key phases: Induction therapy, where high-dose chemotherapy is administered before surgery to shrink tumours, and consolidation therapy, which follows surgery to eradicate any remaining cancer cells and prevent relapse. Radiotherapy has an essential function in NB management for high-risk, especially with tumours that cannot be resected entirely. External Beam Radiotherapy (EBRT) treats the residual areas of the tumour, whereas MIBG treatment, a type of targeted radiotherapy with radioiodinated metaiodobenzylguanidine (MIBG), is utilized with high metastatic disease.

Emerging advances in immunotherapy and targeted therapy have revolutionized the management of NB. Dinutuximab, a monoclonal anti-GD2 antibody, increases survival among high-risk patients by targeting NB cells while sparing normal cells. ALK inhibitors like Crizotinib and Lorlatinib are novel targeted treatments for patients with ALK-mutated NB. Retinoic acid differentiation therapy is also utilized to trigger the maturation of NB cells and eliminate recurrence [8].

AI and ML are revolutionizing NB diagnosis and treatment planning. AI-based models combine tumour genomics, imaging biomarkers, and clinical information to enhance risk stratification, outcome prediction, and personalized therapy choice. AI-aided radiomics improves MRI and CT image analysis, enabling more accurate tumour detection and classification. DL-based histopathological models are also being investigated to detect molecular subtypes and predict therapeutic response, thereby propelling the field of precision oncology [8, 9].

Figure 3. Shimada classification standard [6]

ML has emerged as a transformative tool in NB analysis, enabling computational evaluation of disease progression, therapeutic response, and survival outcomes. The integration of image processing, artificial neural networks (ANNs), multilayer perceptrons (MLPs), support vector machines (SVMs), deep learning models, decision trees, and random forest classifiers has substantially advanced diagnostic accuracy and risk stratification [10, 23]. These algorithms enable automatic feature extraction and classification of histopathological and radiological data, thereby facilitating robust risk prediction.

Figure 4 illustrates the synergistic interaction among these models. The process typically begins with image preprocessing techniques that enhance raw histological and radiological inputs for computational analysis. ANNs and MLPs are used to identify intricate patterns within these data, informing prognosis and treatment pathways [20]. Extracted features are then analyzed using SVMs and random forest classifiers to classify NB into risk categories based on histological grade and imaging biomarkers [12].

Deep learning models extend this capability by integrating transcriptomic, histological, and clinical staging information (such as the INSS and MYCN amplification status), offering deeper insights into tumour biology and enabling patient stratification [15]. Decision tree-based models provide interpretable probabilistic predictions of survival outcomes, aiding oncologists in clinical decision-making. More recently, ensemble methods combining ML and DL frameworks have been shown to improve prediction accuracy by leveraging the strengths of multiple algorithms simultaneously [29].

Despite these achievements, substantial methodological limitations hinder the generalizability and clinical translation of current approaches. Many reviewed studies are based on datasets with limited sample sizes or single-institution sourcing, which restrict population diversity and elevate the risk of overfitting. For instance, studies using fewer than 100 patient samples or isolated histopathological slides often report high accuracy, but these findings may not hold when tested on external datasets. This restricts their use in real-world settings where demographic variability is substantial.

In addition, dataset heterogeneity is a major barrier to reproducibility. Imaging studies vary significantly in magnification levels (e.g., 40× vs. 100×), staining protocols, and scanner calibration. Likewise, omics-based studies differ in gene normalization techniques, sequencing platforms, and preprocessing strategies. Without standardized pipelines, comparisons across models remain confounded by technical artifacts rather than true algorithmic performance. Annotation inconsistencies, particularly in differentiating poorly differentiated NB or stroma-poor subtypes, further impact the reliability of training labels and supervised learning outputs. Another critical concern is the lack of external validation. Most studies rely on random train-test splits or k-fold cross-validation within the same dataset, which does not simulate real-world variability. True external validation using independent cohorts—across institutions and acquisition protocols—is rarely performed. As a result, models trained on homogenous data frequently underperform when exposed to broader populations. This issue calls for more rigorous benchmarking using public datasets and community-driven ML challenges.

To address these limitations, we advocate for the integration of multi-institutional and demographically diverse datasets in future model training. Standardization of preprocessing workflows—such as using stain normalization methods (e.g., Macenko normalization for H&E slides) and unified gene normalization techniques—would reduce variability and improve reproducibility. Model evaluation protocols must incorporate nested cross-validation and independent test sets. Furthermore, open benchmarking platforms should be developed to facilitate fair and transparent algorithm comparisons.

Parallel to these limitations, interpretability remains a significant barrier to clinical adoption. The "black-box" nature of deep learning models hinders transparency and trust. Clinicians are often reluctant to rely on predictions that lack explanatory depth. To counter this, explainable AI (XAI) techniques such as Shapley Additive Explanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM) have been introduced to visualize and interpret model decisions [21, 40, 41]. SHAP provides feature-level contribution scores, while Grad-CAM enables localization of salient image regions influencing decisions. Such tools can bridge the gap between algorithmic output and clinical reasoning [42, 43].

Ethical considerations are equally paramount in paediatric oncology. Ensuring patient confidentiality in training data—especially under frameworks such as the General Data Protection Regulation (GDPR) and Health Insurance Portability and Accountability Act (HIPAA)—is non-negotiable. In addition, biases in training data arising from underrepresentation of specific ethnic or socioeconomic groups can lead to inequitable outcomes [25]. Addressing this requires both adversarial debiasing strategies and deliberate inclusion of diverse cohorts during data acquisition.

Beyond algorithmic accuracy and interpretability, ethical considerations and data fairness are critical in deploying ML models for NB care. One prominent concern is training data bias, particularly when datasets are derived from a single geographic or institutional cohort. Such models may fail to generalize across populations with different genetic, socioeconomic, or ethnic profiles, leading to unintended disparities in diagnostic performance. For instance, underrepresentation of certain risk groups or age demographics could lead to overfitting in well-represented cohorts while underperforming for others [20, 25].

To mitigate this, several strategies have emerged. The inclusion of demographically diverse, multi-institutional datasets is fundamental to reduce institutional bias. Additionally, adversarial debiasing techniques—which train the model to reduce correlation with sensitive attributes (e.g., age, ethnicity, institution)—have shown promise in ensuring equitable predictions [44, 45]. Another emerging practice is the stratified sampling of training and validation cohorts, ensuring balanced representation during both model development and evaluation phases.

Furthermore, ethical compliance with data protection standards such as the GDPR and HIPAA is essential. FL, as discussed later, offers a privacy-preserving alternative by training models across multiple institutions without centralizing data. For equitable deployment, future frameworks must also include bias auditing, ensuring models are assessed across subpopulations and flagged when disproportionate error rates are observed.

One promising solution is the development of FL frameworks, which allow decentralized training across multiple institutions without the need to pool sensitive patient data. This approach not only preserves privacy but also enhances model robustness by incorporating heterogeneous sources. Future research should also focus on integrating multi-modal models that combine histopathology, radiomics, genomics, and clinical records for comprehensive NB risk estimation.

Recent advancements in FL have extended beyond conventional FL algorithms. For instance, Swarm Learning integrates blockchain with federated architectures to coordinate learning across untrusted medical institutions, ensuring both data privacy and verifiable contributions [33]. Similarly, SplitFed merges split learning and FL, enabling scalable training across resource-limited nodes by sharing only intermediary activations rather than raw data [34]. Such frameworks are especially suited to paediatric NB, where institutional silos and ethical restrictions make centralized data pooling unfeasible. Real-world validation, such as FL deployment across 20 hospitals for COVID-19 outcomes [44], demonstrates the viability of such methods in clinical environments. In addition to imaging and histopathological data, the integration of multi-omics layers (e.g., transcriptomics, methylation, proteomics) using deep learning has proven effective in refining risk prediction. A recent study by Dayan et al. [35] utilized a deep survival model on integrated omics datasets to identify complex molecular subtypes in NB, achieving superior predictive performance compared to unimodal models. These strategies support a shift toward biologically grounded stratification and could guide personalized therapeutic interventions by identifying latent prognostic features not visible in single-data domains. Histopathology analysis has also evolved with the application of ViTs, which surpass CNNs in modeling long-range tissue dependencies via self-attention mechanisms. ViTs have demonstrated improved classification accuracy in whole-slide image tasks across oncology domains [31]. To complement this, transformer-specific interpretability tools, such as Layer-wise Relevance Propagation (LRP) and attention rollout methods, enhance transparency by attributing decisions to spatially resolved features [37]. When combined with counterfactual explainability [36], these methods can offer actionable clinical insights while maintaining model trust and compliance with interpretability mandates.

Finally, prospective clinical validation is a critical next step. While most existing studies rely on retrospective data, the true utility of ML models lies in their performance in real-time clinical workflows. Future research must prioritize clinical trials and longitudinal evaluations of AI-driven models in real hospital settings.

While retrospective studies provide a foundational understanding of ML performance in NB diagnosis and prognosis, real-world clinical validation remains critically underexplored. Most of the cited works rely on previously collected datasets—often highly curated—lacking the complexity, noise, and variability of live clinical workflows. Prospective clinical trials are necessary to evaluate how these models perform in routine diagnostics, under diverse patient conditions, institutional workflows, and imaging modalities [11, 13].