مقالات پذیرفته شده در نهمین کنگره بین المللی زیست پزشکی
A Systematic Review of U-Net Based Deep Learning for Glioblastoma Segmentation in Brain MRI
A Systematic Review of U-Net Based Deep Learning for Glioblastoma Segmentation in Brain MRI
Parand Soliemanifard,1Alireza Pourrahim,2,*
1. Student Research Committee, Faculty of Nursing, Dezful University of Medical Sciences, Dezful, Iran. 2. Student Research Committee, Faculty of Medicine, Ilam University of Medical Sciences, Ilam, Iran
Introduction: Accurate segmentation of glioblastoma subregions in brain MRI is critical for prognosis, treatment planning, and predicting overall survival (OS). U-Net and its variants have become the de facto architectures for medical image segmentation, yet their comparative performance, data requirements, and extensions for survival prediction remain unclear.
Methods: We performed a systematic search of PubMed, Embase, IEEE Xplore, Scopus, and Web of Science for studies published from January 2015 to September 2025. Search terms included “U-Net”, “glioblastoma segmentation”, “brain MRI”, and “deep learning”. Two reviewers screened titles, abstracts, and full texts to identify original research applying U-Net–based architectures for automated glioblastoma segmentation, with or without OS prediction. Data extracted encompassed model variants, dataset sources (e.g., BraTS 2019/2021, TCGA-TCIA), preprocessing steps, architectural modifications, performance metrics (Dice coefficient, IoU, Hausdorff distance), computational efficiency, and OS prediction accuracy.
Results: Across the four included studies, U-Net–based models demonstrated consistently high segmentation performance on glioblastoma and other brain tumor subregions in MRI datasets totaling over 13,000 volumes. Whole-tumor Dice coefficients clustered tightly between 0.84 and 0.93 (mean ≈ 0.89), with tumor core Dice scores ranging from 0.80 to 0.92 (mean ≈ 0.86) and enhancing-region scores from 0.63 to 0.88 (mean ≈ 0.76). Integrating boundary priors (Edge U-Net) or ImageNet-pretrained backbones (ResNet-U-Net, Inception-U-Net, VGG-U-Net) yielded relative Dice improvements of 3–6% over the baseline 2D U-Net. One optimized architecture also reported an Intersection-over-Union of 90.4% and a mean Hausdorff distance of 4.1 mm. Streamlined skip connections and encoder–decoder pathways reduced inference time from an average of 5–9 seconds down to approximately 1.5 seconds per volume. A hybrid radiomic–segmentation study achieved 31% overall-survival prediction accuracy using tumor masks and intensity/shape features. Despite strong overall metrics, moderate heterogeneity arose from variable MRI modalities, preprocessing pipelines, and validation schemes, and external validation was limited to a single study.
Conclusion: U-Net–based models consistently deliver high precision in glioblastoma segmentation, with enhancements such as boundary integration, backbone pre-training, and architectural optimizations yielding Dice scores up to 92.5% and real-time inference. Radiomics combined with segmentation can provide preliminary OS estimates, though predictive accuracy remains modest. Heterogeneity in datasets, evaluation metrics, and preprocessing pipelines limits direct comparison. Future work should prioritize multi-institutional validation, standardized reporting (e.g., STAPLE for segmentation benchmarks), and integration of survival modeling into end-to-end deep learning frameworks to translate these advances into clinical practice.