Modality redundancy for MRI-based glioblastoma segmentation.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Automated glioblastoma segmentation typically uses four MRI modalities: T1, contrast-enhanced T1 (T1CE), T2, and FLAIR.
• Redundancy among these modalities may decrease model performance and increase risks associated with missing data in clinical settings.

Purpose of the Study:

• To investigate the relevance and impact of different MRI modalities on glioblastoma segmentation accuracy.
• To determine if reduced input sets can achieve comparable segmentation performance to the standard four-modality input.

Main Methods:

• Trained multiple segmentation models using nnU-Net and SwinUNETR architectures with varying combinations of MRI input modalities.
• Evaluated model performance based on segmentation accuracy and epistemic uncertainty.

Main Results:

• Segmentation using T1CE or T1CE-FLAIR achieved accuracies comparable to the full four-modality input for specific tumor regions.
• nnU-Net showed peak accuracy with T1CE-FLAIR-T1, indicating potential redundancy issues with more inputs.
• SwinUNETR demonstrated statistically equivalent results between three-input and full-input models.

Conclusions:

• A minimal-input model using T1CE-FLAIR is a viable alternative for glioblastoma segmentation.
• Adding modalities beyond T1CE-FLAIR does not consistently improve accuracy and may degrade it, though it can reduce segmentation uncertainty.