Scholarlink

<p>Accurate segmentation of brain tumors in MRI remains challenging due to heterogeneity in appearance, irregular shapes, and low contrast. To address these challenges, this study proposes a novel hybrid framework that integrates a VGG19-based deep encoder feature extractor and the optimization strength of an Improved Spatially Constrained Fish School Optimization (ISCFSO) algorithm to enhance brain tumor segmentation. The VGG19 model is used as an encoder backbone to extracts multi-scale semantic features from magnetic resonance imaging scans, which are then passed through a custom decoder to generate initial segmentation masks. These masks are iteratively refined using a biologically inspired Spatially Constrained Fish School Optimization algorithm. The proposed model which includes elite selection, barycenter-guided movement, simulated annealing-based jump strategies, and dynamic step size adjustments to prevent premature convergence and maintain solution diversity. The framework was tested on the publicly available Figshare brain MRI dataset, and the model outperformed multiple current approaches, with an average segmentation accuracy of 0.9937. This technique increases tumor border definition and resilience over a wide range of situations, potentially improving clinical diagnosis and therapy planning.</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

VGG19 encoder-ISCFSO: An enhanced evolutionary framework for brain tumor segmentation

  • 期刊论文

  • 发表日期：

    2026年6月17日

查看全文
  • Muthulakshmi Kirubakaran,
  • Jayalakshmi Mohan

摘要

Accurate segmentation of brain tumors in MRI remains challenging due to heterogeneity in appearance, irregular shapes, and low contrast. To address these challenges, this study proposes a novel hybrid framework that integrates a VGG19-based deep encoder feature extractor and the optimization strength of an Improved Spatially Constrained Fish School Optimization (ISCFSO) algorithm to enhance brain tumor segmentation. The VGG19 model is used as an encoder backbone to extracts multi-scale semantic features from magnetic resonance imaging scans, which are then passed through a custom decoder to generate initial segmentation masks. These masks are iteratively refined using a biologically inspired Spatially Constrained Fish School Optimization algorithm. The proposed model which includes elite selection, barycenter-guided movement, simulated annealing-based jump strategies, and dynamic step size adjustments to prevent premature convergence and maintain solution diversity. The framework was tested on the publicly available Figshare brain MRI dataset, and the model outperformed multiple current approaches, with an average segmentation accuracy of 0.9937. This technique increases tumor border definition and resilience over a wide range of situations, potentially improving clinical diagnosis and therapy planning.