MMA-LTS: spatially adaptive confidence calibration for missing-modality brain MRI

Lee et al. introduce Missing Modality-Aware Local Temperature Scaling, calibrating voxel-wise brain tumor segmentation confidences when MRI acquisition sequences are incomplete.

Primary source: arXiv:2610.11419: https://arxiv.org/abs/2610.11419

What’s new: Multi-parametric magnetic resonance imaging (mpMRI) for brain tumor segmentation depends on four sequences: T1, contrast-enhanced T1 (T1ce), T2, and FLAIR. In urgent hospital admissions or pediatric scans, scan protocols are often interrupted or incomplete. Although contemporary deep learning models maintain acceptable segmentation overlap under missing modalities, their confidence scores become severely miscalibrated. In Missing Modality-Aware Calibration for Trustworthy Brain Tumor Segmentation (arXiv:2610.11419, 8 Oct 2026), Lee et al. address this safety bottleneck.

Conventional post-hoc calibration methods assume uncertainty scales monotonically with the number of missing sequences. The authors demonstrate that calibration errors depend primarily on which specific modalities are missing and vary non-uniformly across anatomical tumor subregions. To correct this, they develop MMA-LTS (Missing Modality-Aware Local Temperature Scaling), a post-hoc calibration framework that estimates a spatially adaptive, voxel-wise temperature field. The temperature map is conditioned on a learnable modality-availability token and a voxel-level prediction difficulty score. Evaluated on the BraTS 2020 and FeTS 2024 benchmarks across incomplete input combinations, MMA-LTS significantly reduced expected calibration error across diverse missing-modality subsets while fully preserving base model segmentation accuracy.

Why it matters: Neuro-oncology interventions, including surgical resection margins and stereotactic radiosurgery planning, require reliable uncertainty estimates. Uncalibrated overconfidence on incomplete scans risks overtreating healthy brain tissue or missing infiltrating tumor margins. Calibrating voxel-level confidences conditioned on scan completeness provides clinicians with trustworthy indicators of when manual review or repeat imaging is necessary.

Caveats: The framework was developed and evaluated on retrospective curated research cohorts (BraTS 2020 and federated FeTS 2024) under simulated sequence dropout. Prospective clinical validation in emergency departments, where imaging artifacts, non-standard contrast timings, and hardware heterogeneity co-occur with missing scans, has not yet been demonstrated.