Model Ensemble for Brain Tumor Segmentation in Magnetic Resonance Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Accurate brain tumor segmentation in multi-parametric MRI is crucial for quantitative analysis, clinical trials, and personalized patient care.
• This analysis aids in clinical decision-making for diagnosis and prognosis.
• The 2023 Brain Tumor Segmentation (BraTS) challenge expanded to eight tasks with 4,500 cases.

Purpose of the Study:

• To develop and evaluate a deep learning ensemble strategy for segmenting newly included tumor types in the BraTS challenge.
• To assess the performance on pediatric brain tumors (PED), intracranial meningioma (MEN), and brain metastases (MET).

Main Methods:

• An ensemble strategy combining state-of-the-art nnU-Net and Swin UNETR models on a region-wise basis.
• Implementation of a targeted post-processing strategy using cross-validated threshold search for improved tumor sub-region segmentation.
• Evaluation on unseen test cases for PED, MEN, and MET tasks.

Main Results:

• Achieved lesion-wise Dice scores for PED: 0.653 (enhancing tumor), 0.809 (tumor core), 0.826 (whole tumor).
• Achieved lesion-wise Dice scores for MEN: 0.876 (enhancing tumor), 0.867 (tumor core), 0.849 (whole tumor).
• Achieved lesion-wise Dice scores for MET: 0.555 (enhancing tumor), 0.6 (tumor core), 0.58 (whole tumor).
• Ranked first for PED, third for MEN, and fourth for MET in the BraTS challenge.

Conclusions:

• The proposed deep learning ensemble strategy effectively segments pediatric brain tumors, meningiomas, and brain metastases.
• The method demonstrates competitive performance in the challenging BraTS 2023 tasks, ranking highly.
• This approach holds potential for advancing quantitative analysis and supporting clinical decision-making in neuro-oncology.