# Diagnostic Performance of Neural Network-Based Artificial Intelligence in the Detection and Classification of Pediatric Astrocytoma: A Systematic Review

## TL;DR

This paper reviews how AI improves diagnosis of pediatric astrocytoma, showing high accuracy but needing more validation.

## Contribution

Systematic evaluation of neural network-based AI's diagnostic performance in pediatric astrocytoma detection and classification.

## Key findings

- AI models achieved AUROC values exceeding 0.99 and classification accuracies up to 95%.
- AI outperformed radiologists in distinguishing histologically similar tumors.
- Innovations like 3D texture analysis and multi-parametric MRI improved AI performance.

## Abstract

Pediatric posterior fossa tumors, particularly astrocytomas, pose significant diagnostic challenges due to their histological diversity and overlapping imaging features. Conventional methods relying on radiologists' interpretations are prone to subjectivity and variability. Neural network-based artificial intelligence (AI) has emerged as a promising tool to enhance diagnostic accuracy. This systematic review evaluates the diagnostic performance of AI models in pediatric astrocytoma detection and classification, synthesizing evidence from six studies to assess strengths, limitations, and clinical applicability. Following Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines, a systematic search of PubMed, Excerpta Medica database (Embase), Web of Science, and Scopus identified 291 records. After deduplication, 102 studies underwent title/abstract screening, 26 advanced to full-text review, and six met inclusion criteria. Studies were included if they evaluated neural network-based AI for pediatric posterior fossa tumor diagnosis and reported quantitative performance metrics. Risk of bias was assessed using Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2), and findings were synthesized narratively due to methodological heterogeneity. Included studies demonstrated high diagnostic accuracy for AI models, with area under the receiver operating characteristic curve (AUROC) values exceeding 0.99 and classification accuracies up to 95%. AI frequently outperformed radiologists, particularly in distinguishing histologically similar tumors. Innovations like 3D texture analysis and multi-parametric Magnetic resonance imaging (MRI) integration enhanced performance. However, the small number of included studies, limited sample sizes, and retrospective designs limit the generalizability of these findings. Methodological concerns, such as high risk of bias in patient selection and use of subjective reference standards, were also noted. Neural network-based AI shows transformative potential in pediatric astrocytoma diagnostics, offering superior accuracy and efficiency compared to conventional methods. Nonetheless, clinical translation requires addressing these methodological limitations, along with enhancing dataset diversity, ensuring prospective validation, and considering ethical implications. Future research should prioritize multi-center trials, explainable AI frameworks, and integration of multi-modal data to bridge the gap between experimental models and real-world clinical practice.