Scholarlink

<p>Brain Tumor Segmentation (BTS) refers to the automated identification and delineation of tumor regions in brain MRI scans. Deep Learning (DL) has significantly advanced BTS, particularly through multi-modal (MM) imaging. These improvements have enhanced diagnostic precision and support more informed clinical decision-making. This review examined 220 publications from 2017 to 2025, selected for their relevance. The scope covers architectural trends, segmentation challenges, dataset usage, and clinical applicability. Benchmark datasets such as BraTS, TCGA-GBM, and Figshare are commonly used due to their annotated, multi-institutional MRI scans that reflect real-world tumor variability. DL models including CNNs, U-Net variants, and hybrid architectures have shown distinct strengths. Convolutional Neural Networks (CNNs) are effective for hierarchical feature extraction, U-Net models improve localization through skip connections, and hybrid approaches combine attention mechanisms and multi-scale learning to achieve higher accuracy and generalizability. A trend toward more advanced architectures, particularly transformer-based models and self-supervised learning techniques, has emerged in recent years. With sustained improvements in performance and interpretability, DL-based BTS methods hold strong promise for widespread clinical integration, ultimately supporting more accurate, timely, and personalized treatment planning. According to this study, data-related issues such as class imbalance, low contrast, and annotation bias; model-level limitations like morphological ambiguity and insufficient spatiotemporal context; and clinical barriers limit interpretability and deployment feasibility of AI based systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Systematic Review of Deep Learning Approaches for Brain Tumor Segmentation in MRI: Trends, Challenges and Future Directions

  • 期刊论文

  • 发表日期：

    2025年9月30日

查看全文
  • Sandeep Kaur,
  • Usha Mittal,
  • Ankita Wadhawan

摘要

Brain Tumor Segmentation (BTS) refers to the automated identification and delineation of tumor regions in brain MRI scans. Deep Learning (DL) has significantly advanced BTS, particularly through multi-modal (MM) imaging. These improvements have enhanced diagnostic precision and support more informed clinical decision-making. This review examined 220 publications from 2017 to 2025, selected for their relevance. The scope covers architectural trends, segmentation challenges, dataset usage, and clinical applicability. Benchmark datasets such as BraTS, TCGA-GBM, and Figshare are commonly used due to their annotated, multi-institutional MRI scans that reflect real-world tumor variability. DL models including CNNs, U-Net variants, and hybrid architectures have shown distinct strengths. Convolutional Neural Networks (CNNs) are effective for hierarchical feature extraction, U-Net models improve localization through skip connections, and hybrid approaches combine attention mechanisms and multi-scale learning to achieve higher accuracy and generalizability. A trend toward more advanced architectures, particularly transformer-based models and self-supervised learning techniques, has emerged in recent years. With sustained improvements in performance and interpretability, DL-based BTS methods hold strong promise for widespread clinical integration, ultimately supporting more accurate, timely, and personalized treatment planning. According to this study, data-related issues such as class imbalance, low contrast, and annotation bias; model-level limitations like morphological ambiguity and insufficient spatiotemporal context; and clinical barriers limit interpretability and deployment feasibility of AI based systems.