MRI-Based End-To-End Pediatric Low-Grade Glioma Segmentation and Classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Radiomics models using MRI can predict genetic markers in pediatric low-grade glioma (pLGG).
• Manual tumor segmentation is a time-consuming bottleneck for these models.
• Automating segmentation is crucial for efficient pLGG classification.

Purpose of the Study:

• To develop and evaluate an automated deep learning (DL) pipeline for pLGG segmentation and classification.
• To assess the performance of an end-to-end radiomics-based approach for predicting genetic markers in pLGG.
• To compare the efficacy of automated segmentation versus manual segmentation in radiomics models.

Main Methods:

• A two-step U-Net based deep learning network was designed for automated tumor segmentation.
• The first U-Net located tumors in downsampled images; the second refined segmentation using image patches.
• The automated segmentation was integrated into a radiomics model for genetic marker prediction.

Main Results:

• The segmentation model achieved over 80% correlation for volume-related features and a Dice score of .795.
• The radiomics model using auto-segmentation yielded AUCs of .843 (2-class) and .730 (3-class).
• Performance with automated segmentation was comparable to models using manual segmentations (AUCs .874 and .758, respectively).

Conclusions:

• The proposed end-to-end DL pipeline effectively automates pLGG segmentation.
• Automated segmentation integrated with radiomics provides accurate genetic marker prediction comparable to manual methods.
• This approach offers a more efficient and scalable solution for pLGG classification.