Confidence-Aware Brain Tumor Segmentation under Missing MRI Modalities

Brain tumor segmentation from multi-parametric MRI (mpMRI) supports diagnosis, radiotherapy planning, and treatment response monitoring, but most deep learning models assume that all magnetic resonance imaging (MRI) sequences are always available. In real clinical workflows, one or more modalities (e.g., post-contrast T1ce) may be missing due to contraindications, cost, or protocol variations. This paper presents a practical, confidence-aware pipeline for robust tumor segmentation under missing MRI modalities using self-configuring U-Net framework (nnU-Net) trained on the BraTS 2021 dataset. We (i) train a segmentation backbone using all modalities, (ii) train four modality-synthesis models (T1, T1ce, T2, FLAIR) to reconstruct missing sequences, and (iii) apply post-hoc confidence calibration (temperature scaling) to generate voxel-wise confidence maps and a case-level risk score for failure flagging. Our synthesis models achieve stable convergence with validation loss in the range of 0.051–0.063 and peak signal-to-noise ratio (PSNR) around 15–16.7 dB, enabling consistent modality completion. We provide a failure-detection strategy based on calibrated confidence and reliability metrics (expected calibration error (ECE), risk-coverage), designed for safe deployment where uncertain outputs are escalated to human review. The proposed framework targets realistic hospital constraints while maintaining interpretability and quality control.