Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Model Ensemble for Brain Tumor Segmentation in Magnetic Resonance Imaging

Daniel Capellán-Martín, Zhifan Jiang, Abhijeet Parida, Xinyang Liu, Van Lam, Hareem Nisar, Austin Tapp, Sarah Elsharkawi, Maria J. Ledesma-Carbayo, Syed Muhammad Anwar, Marius George Linguraru

TL;DR

This work tackles brain tumor segmentation in BraTS 2023's newly added PED, MEN, and MET tasks with limited data. It introduces a region-wise ensemble of two state-of-the-art models, nnU-Net and Swin UNETR, combined with a cross-validated threshold-based post-processing pipeline to optimize lesion-wise performance. The approach yields improved lesion-wise Dice scores and ranks highly on unseen test data (PED 1st, MEN 3rd, MET 4th), underscoring the value of model ensembling and task-specific post-processing for rare tumors. The study also discusses limitations due to data scarcity and proposes future directions, including self-supervised pretraining and synthetic data augmentation, to boost generalizability across clinical settings.

Abstract

Segmenting brain tumors in multi-parametric magnetic resonance imaging enables performing quantitative analysis in support of clinical trials and personalized patient care. This analysis provides the potential to impact clinical decision-making processes, including diagnosis and prognosis. In 2023, the well-established Brain Tumor Segmentation (BraTS) challenge presented a substantial expansion with eight tasks and 4,500 brain tumor cases. In this paper, we present a deep learning-based ensemble strategy that is evaluated for newly included tumor cases in three tasks: pediatric brain tumors (PED), intracranial meningioma (MEN), and brain metastases (MET). In particular, we ensemble outputs from state-of-the-art nnU-Net and Swin UNETR models on a region-wise basis. Furthermore, we implemented a targeted post-processing strategy based on a cross-validated threshold search to improve the segmentation results for tumor sub-regions. The evaluation of our proposed method on unseen test cases for the three tasks resulted in lesion-wise Dice scores for PED: 0.653, 0.809, 0.826; MEN: 0.876, 0.867, 0.849; and MET: 0.555, 0.6, 0.58; for the enhancing tumor, tumor core, and whole tumor, respectively. Our method was ranked first for PED, third for MEN, and fourth for MET, respectively.

Model Ensemble for Brain Tumor Segmentation in Magnetic Resonance Imaging

TL;DR

This work tackles brain tumor segmentation in BraTS 2023's newly added PED, MEN, and MET tasks with limited data. It introduces a region-wise ensemble of two state-of-the-art models, nnU-Net and Swin UNETR, combined with a cross-validated threshold-based post-processing pipeline to optimize lesion-wise performance. The approach yields improved lesion-wise Dice scores and ranks highly on unseen test data (PED 1st, MEN 3rd, MET 4th), underscoring the value of model ensembling and task-specific post-processing for rare tumors. The study also discusses limitations due to data scarcity and proposes future directions, including self-supervised pretraining and synthetic data augmentation, to boost generalizability across clinical settings.

Abstract

Segmenting brain tumors in multi-parametric magnetic resonance imaging enables performing quantitative analysis in support of clinical trials and personalized patient care. This analysis provides the potential to impact clinical decision-making processes, including diagnosis and prognosis. In 2023, the well-established Brain Tumor Segmentation (BraTS) challenge presented a substantial expansion with eight tasks and 4,500 brain tumor cases. In this paper, we present a deep learning-based ensemble strategy that is evaluated for newly included tumor cases in three tasks: pediatric brain tumors (PED), intracranial meningioma (MEN), and brain metastases (MET). In particular, we ensemble outputs from state-of-the-art nnU-Net and Swin UNETR models on a region-wise basis. Furthermore, we implemented a targeted post-processing strategy based on a cross-validated threshold search to improve the segmentation results for tumor sub-regions. The evaluation of our proposed method on unseen test cases for the three tasks resulted in lesion-wise Dice scores for PED: 0.653, 0.809, 0.826; MEN: 0.876, 0.867, 0.849; and MET: 0.555, 0.6, 0.58; for the enhancing tumor, tumor core, and whole tumor, respectively. Our method was ranked first for PED, third for MEN, and fourth for MET, respectively.
Paper Structure (12 sections, 6 figures, 2 tables)

This paper contains 12 sections, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Data description
  4. Deep Learning Models and Experimental Setup
  5. nnU-Net
  6. Swin UNETR
  7. Model Ensembling
  8. Post-processing
  9. Results
  10. Discussion
  11. Conclusion
  12. Acknowledgements

Figures (6)

  • Figure 1: Proposed method: model ensembling and post-processing pipeline. Outputs are obtained from two state-of-the-art deep learning models. These outputs are subjected to nonlinear activation functions and ensembling strategies. Finally, the ensembled predictions are subjected to a specifically tailored post-processing step.
  • Figure 2: Training examples in the PED, MEN, and MET tasks (from top to bottom) with the following tumor subregions: enhancing tumor ET (blue), a combination of nonenhancing tumor, cystic component, and necrosis NCR (red), and peritumoral edematous area ED in PED or surrounding nonenhancing FLAIR hyperintensity (SNFH) in MEN and MET (green).
  • Figure 3: Post-processing strategy. The ensemble predictions were first cleaned of small disconnected regions. Then, for the PED task, ET and ED labels were redefined based on ET/WT and ED/WT thresholds, respectively.
  • Figure 4: Threshold search on the cross-validation set for identifying small disconnected regions. LW refers to lesion-wise metrics.
  • Figure 5: Threshold search on the cross-validation set. Abbreviations: ED, Peritumoral edematous; ET, enhancing tumor; WT, whole tumor.
  • ...and 1 more figures