Abstract
Abstract Automated brain tumor segmentation has progressed from handcrafted pipelines and U-Net variants to self-configuring 3D CNNs, transformer hybrids, state-space models, diffusion based segmenters, promptable models and brain-specific foundation models. Yet the field’s central scientific problem has changed: another fractional gain in Dice on a familiar BraTS split provides limited evidence that a system will remain reliable when MRI sequences are missing, scanners and institutions change, lesions are small or multifocal, postoperative anatomy is distorted, or uncertainty is high. This structured critical review synthesizes 150 references with emphasis on 2025–2026 and compares architectures, multimodal fusion strategies, preprocessing, missing-modality learning, self-supervision, foundation-model adaptation, federated learning, domain adaptation, uncertainty quantification and clinical translation. We organize evidence around adult glioma, pediatric tumors, brain metastases, meningioma and longitudinal/postoperative applications; distinguish region overlap from lesion detection and volumetric clinical endpoints; and critically examine leakage, site confounding, hidden test reuse, pretraining contamination, annotation uncertainty and challenge-overfitting. Recent evidence is especially informative: brain MRI foundation models such as BrainIAC and BrainFound improve label efficiency and cross-task transfer, while task specific systems such as nnU-Net and carefully engineered 2.5D/3D networks remain formidable full-data baselines. Missing-modality studies increasingly favor adaptive fusion, distillation and modality dropout over naive zero-filling; uncertainty work is moving toward conformal and evidential approaches that can identify cases requiring expert review; and external studies of metastases and meningiomas demonstrate why lesion-wise, site-stratified and human-in-the-loop evaluation is essential. We propose DATA-SEG, Dataset, Annotation, Train/test integrity, Architecture, Statistics, External robustness and Governance, as a reporting and evaluation framework. The review concludes that the next high-impact phase of brain tumor segmentation should prioritize post-cutoff hidden benchmarks, cross-site robustness, missing-sequence stress tests, calibrated selective segmentation, multi-rater probabilistic ground truth and longitudinal clinical readouts rather than leaderboard optimization alone.