Optimizing Brain Tumor Segmentation with MedNeXt: BraTS 2024 SSA and Pediatrics

Sarim Hashmi, Juan Lugo, Abdelrahman Elsayed, Dinesh Saggurthi, Mohammed Elseiagy, Alikhan Nurkamal, Jaskaran Walia, Fadillah Adamsyah Maani, Mohammad YaqubarXiv 2024 · DOI · arXiv

Introduction

This paper, authored by researchers at the Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), details a winning approach to the BraTS 2024 challenge for two specialized tasks: Sub-Saharan Africa (SSA) and Pediatric brain tumor segmentation. By leveraging the MedNeXt architecture and a novel "schedule-free" optimizer, the team achieved state-of-the-art results on datasets representing significant distribution shifts from standard adult glioma benchmarks.

Addressing the Distribution Shift

Most automated brain tumor segmentation models are trained on datasets sourced from the "Global North" and focused on adult patients. This creates a "distribution shift" problem: models often fail when applied to children, whose brain anatomy differs, or to scans from regions like Sub-Saharan Africa, where MRI technology and quality may vary. Manual segmentation remains the gold standard but is slow and prone to human error. There is a critical need for robust, automated tools that can handle these diverse populations to assist radiologists in low-resource settings and specialized pediatric care.

Methodology

The team adopted MedNeXt as their baseline architecture. MedNeXt is a modernized U-Net that incorporates design elements from ConvNeXt, such as depth-wise convolutions (up to 9x9x9) to emulate large receptive fields and inverted bottlenecks to capture complex features.

MedNeXt macro and block architecture illustrating the encoder-decoder structure and specialized blocks.

Key Innovations

  1. Schedule-free Optimizer: Instead of traditional AdamW with a cosine-annealing learning rate schedule, the researchers utilized a "schedule-free" optimizer. This approach removes the need to pre-define a decay schedule, though it required significant hyperparameter tuning to find the optimal learning rate (0.0027 for SSA).
  2. Strategic Finetuning: For the SSA task, which had a limited training set of 60 cases, the team first pre-trained on 1,470 cases from the BraTS 2023 Adult Glioma dataset. They then finetuned only the final decoder blocks and deep supervision layers on the African dataset.
  3. Advanced Post-processing: The team implemented size-thresholding to remove predicted tumor regions smaller than a certain voxel count (e.g., 100 voxels for ET in pediatrics) to reduce false positives. They also manually tuned binarization thresholds (e.g., 0.7 for ET and TC) to optimize Dice scores.

Results and Discussion

The methodology yielded impressive performance on unseen validation sets. For the BraTS-2024 SSA task, the team achieved an average Dice Score of 0.896 and a Hausdorff Distance (HD95) of 14.682. The BraTS-2024 Pediatrics task resulted in an average Dice Score of 0.830 and an HD95 of 37.508.

Ablation study showing the incremental performance gains from patch size changes, auxiliary training, and threshold optimization.

As shown in the ablation study, the finetuning strategy and threshold adjustments (Model #6) provided the best balance, particularly improving the segmentation of the "Enhancing Tumor" (ET) and "Tumor Core" (TC) regions. The results highlight that "borrowing" knowledge from larger adult datasets is more effective than training from scratch when data is scarce.

Conclusion

While the results are strong, the authors note that the pediatric task remains challenging, evidenced by a significantly higher Hausdorff Distance compared to the SSA task. This suggests that pediatric tumors exhibit higher anatomical variability in developing brains. Future work may involve exploring more sophisticated ensembling techniques or incorporating anatomical priors to further stabilize pediatric segmentation.

Summary updated

Abstract

Identifying key pathological features in brain MRIs is crucial for the long-term survival of glioma patients. However, manual segmentation is time-consuming, requiring expert intervention and is susceptible to human error. Therefore, significant research has been devoted to developing machine learning methods that can accurately segment tumors in 3D multimodal brain MRI scans. Despite their progress, state-of-the-art models are often limited by the data they are trained on, raising concerns about their reliability when applied to diverse populations that may introduce distribution shifts. Such shifts can stem from lower quality MRI technology (e.g., in sub-Saharan Africa) or variations in patient demographics (e.g., children). The BraTS-2024 challenge provides a platform to address these issues. This study presents our methodology for segmenting tumors in the BraTS-2024 SSA and Pediatric Tumors tasks using MedNeXt, comprehensive model ensembling, and thorough postprocessing. Our approach demonstrated strong performance on the unseen validation set, achieving an average Dice Similarity Coefficient (DSC) of 0.896 on the BraTS-2024 SSA dataset and an average DSC of 0.830 on the BraTS Pediatric Tumor dataset. Additionally, our method achieved an average Hausdorff Distance (HD95) of 14.682 on the BraTS-2024 SSA dataset and an average HD95 of 37.508 on the BraTS Pediatric dataset. Our GitHub repository can be accessed here: Project Repository : https://github.com/python-arch/BioMbz-Optimizing-Brain-Tumor-Segmentation-with-MedNeXt-BraTS-2024-SSA-and-Pediatrics
BibTeX
@misc{hashmi2024optimizing,
  author = {Sarim Hashmi and Juan Lugo and Abdelrahman Elsayed and Dinesh Saggurthi and Mohammed Elseiagy and Alikhan Nurkamal and Jaskaran Walia and Fadillah Adamsyah Maani and Mohammad Yaqub},
  title = {{Optimizing Brain Tumor Segmentation with MedNeXt: BraTS 2024 SSA and Pediatrics}},
  year = {2024},
  eprint = {2411.15872},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.15872},
}
Figure