Improving Deep Learning Models for Pediatric Low-Grade Glioma Tumours Molecular Subtype Identification Using...

Area of Science:

• Neuro-oncology
• Medical imaging
• Artificial intelligence in medicine

Background:

• Pediatric low-grade gliomas (pLGG) are the most common pediatric brain tumors.
• Accurate molecular diagnosis of pLGG is crucial for targeted treatment strategies.
• Current diagnostic methods can be improved with advanced imaging and computational techniques.

Purpose of the Study:

• To develop and evaluate MRI-based Convolutional Neural Networks (CNNs) for molecular subtype identification of pLGG.
• To augment CNN models with tumor location probability maps for improved diagnostic performance.
• To compare the efficacy of location-based, CNN-based, and hybrid approaches for pLGG molecular diagnosis.

Main Methods:

• Retrospective analysis of MRI FLAIR sequences from 214 pediatric patients with pLGG.
• Development of three diagnostic pipelines: location-based, CNN-based, and hybrid (tumor-location-guided CNNs).
• Statistical validation using Area Under the Receiver Operating Characteristic Curve (AUROC) and Student's t-test over 100 repetitions.

Main Results:

• The location-based classifier achieved an AUROC of 77.9.
• CNN-based classifiers reached an AUROC of 86.1.
• Tumor-location-guided CNNs demonstrated superior performance with an average AUROC of 88.64 (p=0.0018).

Conclusions:

• Incorporating tumor location probability maps into CNN models significantly improves molecular subtype identification in pLGG.
• This hybrid approach offers a more accurate and reliable method for diagnosing pediatric low-grade gliomas.
• Enhanced diagnostic accuracy facilitates more precise and targeted treatment for pediatric brain tumors.