Scholarlink

<p>The clinical importance of obtaining precise automated brain tumor segmentation from magnetic resonance imaging (MRI) data stands high because it leads to both better diagnostic processes and optimal treatment designs and disease monitoring strategies. Research studies need to establish comprehensive comparisons regarding deep learning model performance benefits and operational stability in this task. The research analyzes the segmentation strength of Universal-Net (U-Net) and Volumetric-Net (V-Net) deep learning architectures together with their advanced versions which combine attention mechanisms like Attention U-Net with residual connection systems like residual network-based Universal-Net (ResNet-based U-Net). This study used brain tumor segmentation (BraTS) 2021 public data for multi-modality MRI scans and professional annotations of tumor areas to conduct an exhaustive assessment. The evaluation utilized cross-validation to train the models while using standard scoring metrics such as Dice score and Hausdorff distance and sensitivity particularly for edema and enhancing tumor and necrotic/non-enhancing tumor category segmentations. The study examined preprocessing effects between normalization and data augmentation methods alongside loss function evaluations on segmentation accuracy results. The research presents an extensive comparative review which demonstrates how each model succeeds and fails at managing natural brain tumor shape diversity during analysis. The research findings enhance our understanding of the best choice of model selection based on specific medical requirements and recommend future research objectives to enhance generalization performance while minimizing computational expenses.</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

Brain tumor segmentation capabilities of 3D deep learning architectures (U-Net, V-Net, Attention U-Net, ResNet-based U-Net, Transformer-based model)

  • 期刊论文

  • 发表日期：

    2025年10月3日

查看全文
  • Otabek Puladjonov,
  • Pooja,
  • Ambuj Aggarwal

摘要

The clinical importance of obtaining precise automated brain tumor segmentation from magnetic resonance imaging (MRI) data stands high because it leads to both better diagnostic processes and optimal treatment designs and disease monitoring strategies. Research studies need to establish comprehensive comparisons regarding deep learning model performance benefits and operational stability in this task. The research analyzes the segmentation strength of Universal-Net (U-Net) and Volumetric-Net (V-Net) deep learning architectures together with their advanced versions which combine attention mechanisms like Attention U-Net with residual connection systems like residual network-based Universal-Net (ResNet-based U-Net). This study used brain tumor segmentation (BraTS) 2021 public data for multi-modality MRI scans and professional annotations of tumor areas to conduct an exhaustive assessment. The evaluation utilized cross-validation to train the models while using standard scoring metrics such as Dice score and Hausdorff distance and sensitivity particularly for edema and enhancing tumor and necrotic/non-enhancing tumor category segmentations. The study examined preprocessing effects between normalization and data augmentation methods alongside loss function evaluations on segmentation accuracy results. The research presents an extensive comparative review which demonstrates how each model succeeds and fails at managing natural brain tumor shape diversity during analysis. The research findings enhance our understanding of the best choice of model selection based on specific medical requirements and recommend future research objectives to enhance generalization performance while minimizing computational expenses.