Task-Efficient Modality Selection for Deep Learning–Based Pre-Operational Glioma Segmentation
Abstract
Purpose
Deep learning–based glioma segmentation models are commonly developed under the assumption that the standard glioma MRI protocol—T1-weighted, contrast-enhanced T1 (T1ce), T2-weighted, and FLAIR—is available; however, this assumption may not hold in time-constrained or incomplete clinical imaging workflows. This study aims to quantitatively identify which MRI modalities or modality combinations are sufficient for accurate segmentation of individual glioma subregions.
Methods
The BraTS 2021 dataset was used, consisting of multi-institutional brain MRI scans of adult glioma patients with four structural modalities (T1, T1ce, T2, and FLAIR) and expert voxel-wise annotations for edema (ED), enhancing tumor (ET), and necrotic core (NCR). A total of 1,251 cases were included, with 210 cases reserved as an independent test set and 1,041 cases used for training and validation. Eight 3D transformer-based nnU-Net models were trained using different single-modality and multi-modality input combinations. Segmentation performance was evaluated on the test set using Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), and average surface distance (ASD).
Results
For ED segmentation, FLAIR achieved the highest single-modality DSC (0.81 ± 0.14). Multi-modality inputs yielded modest improvements, with DSC values of 0.86 ± 0.12 for T1ce+FLAIR and 0.87 ± 0.12 for both T1+T1ce+FLAIR and the four-modality configuration, indicating diminishing improvement beyond FLAIR-containing combinations. For ET segmentation, T1ce produced the highest single-modality DSC (0.83 ± 0.23), while multi-modality models incorporating T1ce achieved comparable performance (0.84–0.85 ± 0.23). Similarly, for NCR segmentation, T1ce yielded the highest single-modality DSC (0.77 ± 0.29), with no substantial gains from additional modalities. Overall, the combination of T1ce and FLAIR achieved performance comparable to the four-modality input across all subregions.
Conclusion
Accurate deep learning–based glioma subregion segmentation does not require the complete four-modality MRI protocol. The combination of T1ce and FLAIR is sufficient for robust segmentation, supporting task-efficient modality selection and clinically realistic AI deployment.