Uncertainty-aware feature mapping and adaptive inference for brain tumor segmentation with missing contrast-enhanced...

Multi-parametric magnetic resonance imaging (mpMRI) is a cornerstone of clinical brain tumor assessment, yet the contrast-enhanced T1-weighted (T1ce) sequence is frequently unavailable due to contrast-agent contraindications or acquisition failures. Existing segmentation methods degrade substantially when key modalities are missing, motivating the need for robust missing-modality segmentation frameworks.

We propose UAF-AIMM (Uncertainty-Aware Feature Mapping and Adaptive Inference for Brain Tumor Segmentation with Missing Contrast-Enhanced T1-Weighted Imaging), a 3D brain tumor segmentation framework for the clinically prevalent missing-T1ce scenario. UAF-AIMM integrates three components: (1) MR-Mapper, a cross-modal feature mapping module that recovers T1ce-related semantics in the latent space from observed sequences; (2) Uncertainty-Aware Attention (UAA), which estimates voxel-wise predictive variance and uses learnable gating to fuse features according to their estimated reliability; and (3) Single-Image Test-Time Adaptation (SITA), a per-case calibration strategy that updates only normalization affine parameters to mitigate domain shift at inference.

Evaluated on the BraTS (Brain Tumor Segmentation) 2021 dataset under the missing-T1ce protocol, UAF-AIMM achieves an average Dice similarity coefficient of 89.20% and a 95th percentile Hausdorff distance (HD95) of 5.50 mm, outperforming representative baselines including 3D U-Net, Swin UNETR, mmFormer, and ReCoSeg. The improvement is most pronounced in the contrast-dependent enhancing tumor (ET) subregion.

The results demonstrate that feature-space compensation with explicit reliability control and test-time calibration provides a practical and effective direction for robust brain tumor segmentation under incomplete multi-modal MRI protocols, with potential for translation to heterogeneous clinical environments.