[Paper Review] BraTS-PEDs: Results of the Multi-Consortium International Pediatric Brain Tumor Segmentation Challenge 2023

Anahita Fathi Kazerooni, Nastaran Khalili+78 more|arXiv (Cornell University)|Jul 11, 2024

Radiomics and Machine Learning in Medical Imaging5 citations

TL;DR

This paper presents BraTS-PEDs 2023, the first pediatric-focused BraTS challenge, detailing dataset, evaluation protocol, participating methods, and benchmarked segmentation performance on pediatric high-grade gliomas.

ABSTRACT

Pediatric central nervous system tumors are the leading cause of cancer-related deaths in children. The five-year survival rate for high-grade glioma in children is less than 20%. The development of new treatments is dependent upon multi-institutional collaborative clinical trials requiring reproducible and accurate centralized response assessment. We present the results of the BraTS-PEDs 2023 challenge, the first Brain Tumor Segmentation (BraTS) challenge focused on pediatric brain tumors. This challenge utilized data acquired from multiple international consortia dedicated to pediatric neuro-oncology and clinical trials. BraTS-PEDs 2023 aimed to evaluate volumetric segmentation algorithms for pediatric brain gliomas from magnetic resonance imaging using standardized quantitative performance evaluation metrics employed across the BraTS 2023 challenges. The top-performing AI approaches for pediatric tumor analysis included ensembles of nnU-Net and Swin UNETR, Auto3DSeg, or nnU-Net with a self-supervised framework. The BraTSPEDs 2023 challenge fostered collaboration between clinicians (neuro-oncologists, neuroradiologists) and AI/imaging scientists, promoting faster data sharing and the development of automated volumetric analysis techniques. These advancements could significantly benefit clinical trials and improve the care of children with brain tumors.

Motivation & Objective

• Benchmark and standardize automated volumetric segmentation of pediatric brain tumors from multi-institution mpMRI data.
• Provide ground-truth pediatric tumor subregion annotations and a fair, multi-center evaluation framework.
• Compare diverse AI approaches (primarily U-Net based) for WT, TC, and ET segmentation on BraTS-PEDs data.

Proposed method

• Curated multi-institution pediatric MRI dataset with four sequences (T1, T1CE, T2, FLAIR) and ground-truth labels for ET, NC, CC, and ED (mapped to three labels for challenge).
• Pre-processing to 1 mm3 resolution, co-registration to template, and de-identification using pediatric skull-stripping.
• Participants trained on provided data; evaluation used lesionwise Dice and HD95 across ET, TC, WT with 24 test subjects.
• Ground truth segmentations refined by neuroradiologists and ASNR reviewers following standardized annotation guidelines.

Figure 1: A schematic diagram of different steps in data preparation for BraTS-PEDs 2023 data: (A) Data preparation process; (B) Tumor segmentation approach: Left - original annotations including four subregions, Right – final annotations provided to the teams for model training (the final annotatio

Experimental results

Research questions

• RQ1 How well can AI segmentation methods reproduce pediatric brain tumor subregions (ET, NC/TC, WT) on multi-institution mpMRI data?
• RQ2 Do ensembles of state-of-the-art architectures (e.g., nnU-Net, Swin UNETR) outperform single-model approaches in pediatric tumor segmentation?
• RQ3 What is the impact of pediatric-specific ground-truth annotation and preprocessing on cross-institution segmentation performance?
• RQ4 Is there statistically significant performance difference between top teams on the BraTS-PEDs 2023 test set?
• RQ5 Can ensembling and self-supervised pretraining improve robustness across heterogeneous pediatric MRI data?

Key findings

• Top approaches were ensembles of nnU-Net and Swin UNETR, Auto3DSeg, or nnU-Net with self-supervised pretraining.
• Mean Dice for WT around 0.83–0.84 and TC around 0.77–0.81 across top teams.
• ET segmentation showed more variability with Dice ~0.55–0.65, reflecting challenges with small or absent ET regions in DMG/DIPG.
• Ensembling may improve overall performance by combining strengths and mitigating weaknesses.
• Statistical significance of top-team differences was limited by small test cohort (n=24).
• WT segmentation was the most reliable across teams; ET remained the most challenging subregion.

Figure 2: Pairwise p-values between the participating teams.