Scholarlink

<p>Accurate segmentation of low-grade gliomas (LGGs) remains challenging due to their infiltrative nature and heterogeneous radiological appearance. We present a novel deep learning framework that integrates YOLOv8, Gold-YOLO, and Dynamic Snake Convolution (DSC) for enhanced LGG segmentation in MRI images. Our approach adapts YOLOv8 for semantic segmentation and incorporates DSC, an adaptive attention mechanism that dynamically adjusts convolutional kernels based on local context to capture irregular tumor morphologies. The YOLOv8-DSC architecture effectively delineates tumor boundaries while suppressing imaging artifacts. Training employs a hybrid loss function combining Binary Cross-Entropy (BCE), Dice, and Focal losses to optimize pixel-level accuracy, geometric similarity, and class balance. Gold-YOLO integration as a teacher network produces the final YOLOv8-DSC-Gold model, excelling at small lesion detection while maintaining computational efficiency. Evaluation on the Kaggle LGG Segmentation Dataset, containing diverse tumor characteristics, demonstrates substantial improvements over baseline models. Compared to YOLOv5, YOLOv7, YOLOv8, and YOLOv11, our architecture achieves segmentation Mean Average Precision (mAP50) improvements of 17.7%, 9.8%, 5.1%, and 0.8%, and Dice coefficient gains of 18.9%, 8.0%, 6.4%, and 0.5%, respectively. These results establish YOLOv8-DSC-Gold as a robust solution for automated LGG segmentation in clinical neuroimaging applications.</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

Deep fusion of YOLOv8, gold yolo, and dynamic snake convolution for accurate low-grade glioma segmentation

  • 期刊论文

  • 发表日期：

    2026年6月30日

查看全文
  • Zeinab F. Elsharkawy

摘要

Accurate segmentation of low-grade gliomas (LGGs) remains challenging due to their infiltrative nature and heterogeneous radiological appearance. We present a novel deep learning framework that integrates YOLOv8, Gold-YOLO, and Dynamic Snake Convolution (DSC) for enhanced LGG segmentation in MRI images. Our approach adapts YOLOv8 for semantic segmentation and incorporates DSC, an adaptive attention mechanism that dynamically adjusts convolutional kernels based on local context to capture irregular tumor morphologies. The YOLOv8-DSC architecture effectively delineates tumor boundaries while suppressing imaging artifacts. Training employs a hybrid loss function combining Binary Cross-Entropy (BCE), Dice, and Focal losses to optimize pixel-level accuracy, geometric similarity, and class balance. Gold-YOLO integration as a teacher network produces the final YOLOv8-DSC-Gold model, excelling at small lesion detection while maintaining computational efficiency. Evaluation on the Kaggle LGG Segmentation Dataset, containing diverse tumor characteristics, demonstrates substantial improvements over baseline models. Compared to YOLOv5, YOLOv7, YOLOv8, and YOLOv11, our architecture achieves segmentation Mean Average Precision (mAP50) improvements of 17.7%, 9.8%, 5.1%, and 0.8%, and Dice coefficient gains of 18.9%, 8.0%, 6.4%, and 0.5%, respectively. These results establish YOLOv8-DSC-Gold as a robust solution for automated LGG segmentation in clinical neuroimaging applications.