What question did this study set out to answer?

This review aims to evaluate deep learning techniques for brain tumor segmentation in MRI, focusing on their robustness and clinical translation.

July 8, 2026Open Access

Deep learning for brain tumour segmentation on MRI with focus on robustness uncertainty and clinical translation

Key Points

This review aims to evaluate deep learning techniques for brain tumor segmentation in MRI, focusing on their robustness and clinical translation.
Systematic review following PRISMA 2020 guidelines, synthesizing studies published from January 2018 to March 2025.
Inclusion of 111 studies after screening 2280 records and assessing 292 texts.
Organized findings into five methodological paradigms including CNN architectures and GAN-based techniques.
Transformer-based models showed the highest benchmark Dice scores, indicating better segmentation performance.
Diffusion and probabilistic models provided stronger uncertainty estimation but were less mature for clinical deployment.
Most studies relied on BraTS datasets, with limited external validation and calibration analyses reported.

Abstract

Accurate brain tumour segmentation on magnetic resonance imaging (MRI) is central to diagnosis, radiotherapy planning, radiomic analysis, and longitudinal monitoring. This PRISMA 2020-guided systematic review synthesises deep learning approaches for MRI-based brain tumour segmentation published between January 2018 and March 2025. After screening 2280 records and assessing 292 full texts, 111 studies were included. The review organises the literature into five methodological paradigms: convolutional neural network (CNN)/U-Net architectures, attention and hybrid encoder-decoder models, generative adversarial network (GAN)-based segmentation and augmentation, transformer and CNN-transformer models, and diffusion/probabilistic approaches. Beyond Dice performance, the synthesis examines validation design, benchmark dependence, uncertainty estimation, calibration, explainability, risk of bias, and clinical readiness. Most studies relied exclusively on BraTS datasets, while external validation, calibration analysis, and clinically interpretable uncertainty were uncommon. Transformer-based models generally reported the highest benchmark Dice scores, whereas diffusion and Bayesian approaches offered stronger uncertainty-aware modelling but remain less mature for deployment. The review concludes that future progress requires external multi-centre validation, transparent reporting, calibrated uncertainty, and prospective clinical workflow evaluation beyond benchmark optimisation.

AI에게 질문

Bookmark

View Full Paper