Scholarlink

<p>Brain tumors remain a formidable challenge in clinical neuroscience, necessitating early and accurate detection to enhance treatment efficacy and patient survival. Magnetic Resonance Imaging (MRI) enables high-resolution, non-invasive visualization of brain abnormalities; however, accurate delineation of malignant tumors such as glioblastoma remains difficult due to their complex morphology and diffuse boundaries. Multi-threshold image segmentation is vital for accurately isolating tumor regions, yet conventional techniques often suffer from local optima entrapment and increasing computational complexity as the number of thresholds rises. To overcome these limitations, this study employs a Lévy Flight and Chaos Theory-based Gravitational Search Algorithm (LCGSA) for efficient and robust multi-threshold segmentation of brain MRI images. The algorithm integrates Lévy flight dynamics to enhance global exploration and chaotic maps to refine local exploitation, ensuring a balanced optimization process. Kapur’s entropy is employed as the fitness criterion to determine optimal thresholds for effective region separation. The proposed method is comprehensively evaluated using quantitative metrics, including PSNR, SSIM, FSIM, and MSE, supported by qualitative analyses such as convergence behavior, histograms, and segmented outputs. Statistical validation using Wilcoxon rank-sum and Friedman tests confirms the superiority and consistency of LCGSA over ten competitive metaheuristic algorithms, demonstrating improved segmentation accuracy, robustness, and computational efficiency. Furthermore, to evaluate its adaptability in deep learning-based segmentation, an ablation study is conducted by integrating LCGSA with the U-Net++ architecture and tested on the Kaggle Brain MRI dataset. The hybrid LCGSA–UNet++ model achieves superior performance compared to state-of-the-art deep learning frameworks, yielding a Dice coefficient of approximately 0.92, an IoU of 0.85, and a Hausdorff distance of 4.44, thereby underscoring its efficacy in precise lesion boundary delineation and enhanced generalization for brain tumor segmentation.</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

Lévy flight and chaos-enhanced gravitational search algorithm with U-Net++ for multi-threshold brain MRI segmentation

  • 期刊论文

  • 发表日期：

    2026年2月21日

查看全文
  • Sajad Ahmad Rather,
  • Partha Pratim Roy

摘要

Brain tumors remain a formidable challenge in clinical neuroscience, necessitating early and accurate detection to enhance treatment efficacy and patient survival. Magnetic Resonance Imaging (MRI) enables high-resolution, non-invasive visualization of brain abnormalities; however, accurate delineation of malignant tumors such as glioblastoma remains difficult due to their complex morphology and diffuse boundaries. Multi-threshold image segmentation is vital for accurately isolating tumor regions, yet conventional techniques often suffer from local optima entrapment and increasing computational complexity as the number of thresholds rises. To overcome these limitations, this study employs a Lévy Flight and Chaos Theory-based Gravitational Search Algorithm (LCGSA) for efficient and robust multi-threshold segmentation of brain MRI images. The algorithm integrates Lévy flight dynamics to enhance global exploration and chaotic maps to refine local exploitation, ensuring a balanced optimization process. Kapur’s entropy is employed as the fitness criterion to determine optimal thresholds for effective region separation. The proposed method is comprehensively evaluated using quantitative metrics, including PSNR, SSIM, FSIM, and MSE, supported by qualitative analyses such as convergence behavior, histograms, and segmented outputs. Statistical validation using Wilcoxon rank-sum and Friedman tests confirms the superiority and consistency of LCGSA over ten competitive metaheuristic algorithms, demonstrating improved segmentation accuracy, robustness, and computational efficiency. Furthermore, to evaluate its adaptability in deep learning-based segmentation, an ablation study is conducted by integrating LCGSA with the U-Net++ architecture and tested on the Kaggle Brain MRI dataset. The hybrid LCGSA–UNet++ model achieves superior performance compared to state-of-the-art deep learning frameworks, yielding a Dice coefficient of approximately 0.92, an IoU of 0.85, and a Hausdorff distance of 4.44, thereby underscoring its efficacy in precise lesion boundary delineation and enhanced generalization for brain tumor segmentation.