Auto3DSeg for Brain Tumor Segmentation from 3D MRI in BraTS 2023 Challenge

TL;DR

This work tackles multi-task 3D brain tumor segmentation in BraTS 2023 by applying Auto3DSeg from MONAI to automate data analysis, model configuration, training, and ensembling across five sub-challenges. The authors employ SegResNet within a 5-fold cross-validation framework, with a per-subregion ensembling strategy and a missing-modality adaptation for brain metastasis data, achieving top placements across tasks. Key contributions include automated hyperparameter search, subregion-specific ensembling, and modality-robust training that yield 1st place in three tasks and 2nd place in the other two. The results demonstrate that an automated, scalable 3D segmentation pipeline can achieve state-of-the-art performance across heterogeneous MRI datasets, reducing the manual tuning burden for researchers.

Abstract

In this work, we describe our solution to the BraTS 2023 cluster of challenges using Auto3DSeg from MONAI. We participated in all 5 segmentation challenges, and achieved the 1st place results in three of them: Brain Metastasis, Brain Meningioma, BraTS-Africa challenges, and the 2nd place results in the remaining two: Adult and Pediatic Glioma challenges.

Figures (2)

• Figure 1: A typical segmentation example with true and predicted labels overlaid over T1c MRI axial, sagittal and coronal slices. The whole tumor (WT) class includes all visible labels (a union of green, yellow and red labels), the tumor core (TC) class is a union of red and yellow, and the enhancing tumor core (ET) class is shown in yellow (a hyperactive tumor part).
• Figure 2: SegResNet network configuration. The network uses repeated ResNet blocks with batch normalization and deep supervision