# Development of intelligent tools to predict neuroblastoma risk stratification and overall prognosis based on multiphase enhanced CT and clinical features

## TL;DR

This study creates a deep learning model using CT scans and clinical data to better predict the risk and prognosis of neuroblastoma in children.

## Contribution

A novel deep learning model combining multiphase CT and clinical features for neuroblastoma risk and prognosis prediction.

## Key findings

- Swin-ART achieved 0.770 AUC and 0.780 accuracy for risk stratification in testing.
- The combined Cox and RSF model had a mean C-index of 0.84 for prognosis.
- Multimodal models outperformed single-modality clinical models in accuracy and stability.

## Abstract

Neuroblastoma (NB) is a common malignancy in children, and accurate risk stratification and prognostic assessment are essential for personalized treatment. Current tumor assessment methods rely on clinical features and conventional imaging techniques, which have limited predictive accuracy. The aim of this study was to develop a deep learning model based on multiphase enhanced CT images and clinical features to improve the accuracy of risk stratification and prognostic assessment of NB.

Multi-phase enhanced CT images and clinical features from 202 NB patients were collected. Four risk stratification classifiers were developed using the Swin Transformer model and evaluated in training and testing cohorts. Prognostic models were constructed using a combination of multiple machine learning algorithms in conjunction with CT image features and clinical characteristics.

Swin-ART based on arterial phase images was the best risk stratification classifier with an AUC of 0.770 (95% CI: 0.613–0.909) and an accuracy of 0.780 in the testing cohort. In the prognostic assessment, the combined model of backward stepwise Cox regression and randomized survival forest (RSF) obtained the highest mean C-index of 0.84. The 1-, 3-, and 5-year AUC values of the optimal prognostic model in the training cohort were 0.93 (95% CI: 0.927–0.942), 0.93 (95% CI: 0.929–0.946), and 0.96 (95% CI: 0.953–0.974), respectively. The corresponding AUC values for the testing cohort were 0.90 (95% CI: 0.857–0.934), 0.87 (95% CI: 0.808–0.928), and 0.91 (95% CI: 0.718–0.977), respectively. Multimodal models outperform single-modality clinical models in both predictive accuracy and stability.

This study successfully developed a deep learning model based on multiphase enhanced CT images and clinical features to predict risk stratification and prognosis in NB. The findings provide a new tool for clinical practice and lay the foundation for future precision medicine and personalized treatment.