Scholarlink

<p>Recently, liver tumor segmentation gained significant attention in medical imaging by serving as key step for accurate disease assessment, and precise treatment planning. However, segmentation remained challenging for weakly defined tumors with irregular boundaries that blend in surrounding tissues like muscles and fat, often leading to inaccuracies. To address this issue, Squeeze–Excitation Attention-based Liver Segmentation Network (SEALS-Net) is proposed, by integrating the attention mechanisms to assign precise importance scores to feature maps for improved tumor delineation. The network is trained by using hybrid loss function combining Jaccard and Dice losses, which effectively captures overlapping regions and enhancing segmentation accuracy and then optimized with AdamW optimizer for stable convergence. For classification, Uncertainty based Bayesian Gated Recurrent Unit (UBGRU) model is employed, by utilizing GRU’s ability to capture spatial dependencies in lesion features while Bayesian modeling introduces uncertainty estimation, improving robustness and handling noisy data which are critical for reliable medical decision-making. The experimental results demonstrate the superiority of proposed approach, by achieving a Dice Similarity Coefficient of 96.85% and a Jaccard score of 93.48%, outperforming existing modified SegNet in liver tumor segmentation tasks.</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

SEALS-Net: hybrid loss-guided segmentation and uncertainty based bayesian GRU classification for liver tumor

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Pallavi B.,
  • Meesala Shobha Rani

摘要

Recently, liver tumor segmentation gained significant attention in medical imaging by serving as key step for accurate disease assessment, and precise treatment planning. However, segmentation remained challenging for weakly defined tumors with irregular boundaries that blend in surrounding tissues like muscles and fat, often leading to inaccuracies. To address this issue, Squeeze–Excitation Attention-based Liver Segmentation Network (SEALS-Net) is proposed, by integrating the attention mechanisms to assign precise importance scores to feature maps for improved tumor delineation. The network is trained by using hybrid loss function combining Jaccard and Dice losses, which effectively captures overlapping regions and enhancing segmentation accuracy and then optimized with AdamW optimizer for stable convergence. For classification, Uncertainty based Bayesian Gated Recurrent Unit (UBGRU) model is employed, by utilizing GRU’s ability to capture spatial dependencies in lesion features while Bayesian modeling introduces uncertainty estimation, improving robustness and handling noisy data which are critical for reliable medical decision-making. The experimental results demonstrate the superiority of proposed approach, by achieving a Dice Similarity Coefficient of 96.85% and a Jaccard score of 93.48%, outperforming existing modified SegNet in liver tumor segmentation tasks.