Advanced CT Radiomics and Deep Learning Differentiate Pediatric Neuroblastoma from Ganglioneuroblastoma

Overview

This study developed radiomic and deep learning models based on contrast-enhanced CT combined with clinical and biochemical data to differentiate pediatric peripheral neuroblastoma (NB) from ganglioneuroblastoma (GNB). The models demonstrated improved diagnostic accuracy over conventional imaging, supporting personalized treatment strategies.

Background

Neuroblastoma (NB) is a common extracranial solid tumor in children, originating from sympathetic nervous system cells and presenting a spectrum from malignant NB to benign ganglioneuroma. Differentiating NB from ganglioneuroblastoma (GNB) is clinically challenging due to overlapping clinical and imaging features. Contrast-enhanced CT is essential for tumor evaluation but has limited ability to distinguish NB from GNB. Radiomics and deep learning offer promising tools to enhance diagnostic precision by analyzing imaging features beyond human perception, especially when combined with clinical and biochemical markers.

Data Highlights

A total of 225 pediatric patients (161 NB, 64 GNB) with mean age 2.23 years were retrospectively analyzed. Data were split into training (135 cases) and validation (90 cases) sets. CT scans included arterial and venous phases using standardized protocols. Biochemical markers such as serum ferritin and neuron-specific enolase were collected. A 3D-UNet deep learning model was developed for automated tumor segmentation to facilitate radiomic feature extraction.

Key Findings

• The combined radiomic and deep learning model based on contrast-enhanced CT significantly improved differentiation between NB and GNB compared to conventional imaging alone.
• Automated 3D-UNet segmentation provided consistent and efficient tumor delineation, reducing manual variability.
• Incorporation of clinical data and biochemical indicators (serum ferritin, NSE) enhanced model diagnostic performance.
• Radiomic features captured tumor heterogeneity and microenvironment characteristics not reflected by biochemical markers alone.
• The model supports pretreatment risk stratification and personalized therapeutic decision-making in pediatric NB patients.

Clinical Implications

Integrating advanced radiomic analysis and deep learning with contrast-enhanced CT can aid clinicians in accurately distinguishing malignant neuroblastoma from ganglioneuroblastoma in children. This approach facilitates precise risk assessment and tailored treatment planning, potentially improving outcomes. Automated segmentation streamlines workflow and enhances reproducibility in pediatric tumor imaging.

Conclusion

The study demonstrates that combining CT-based radiomics, deep learning, and clinical-biochemical data provides a robust tool for differentiating pediatric peripheral neuroblastoma from ganglioneuroblastoma. This methodology holds promise for improving diagnostic accuracy and guiding personalized management in pediatric oncology.

References

1. Neuroblastoma and ganglioneuroblastoma clinical context and imaging -- Various sources 2014-2024
2. 3D-UNet architecture for volumetric medical image segmentation -- Literature 2019