Scholarlink

<p>The rising incidence of thyroid cancer highlights the importance of ultrasound imaging for early detection and diagnosis. Accurate segmentation and classification of thyroid nodules in ultrasound images are essential for improving diagnostic precision and guiding clinical treatment. However, conventional deep learning methods often address these tasks separately, missing the potential benefits of tackling them together and limiting overall effectiveness. In this study, we propose a novel multi-task learning (MTL) framework with shared multi-layer parameters to perform nodule segmentation and classification simultaneously. Our architecture includes a Hierarchical Shared Network (HSN) for strong feature extraction, a Task Interaction Module (TIM) to enable dynamic information exchange between tasks, and a Dynamic Decoding Module (DDM) for enhanced segmentation. Experiments on the public TN3K and TG3K datasets demonstrate that our MTL framework achieves a classification accuracy of 93.4% and a segmentation Dice coefficient of 87.8%, outperforming several established approaches. Ablation studies further confirm the positive impact of HSN, TIM, and DDM on overall performance. These findings show that our integrated MTL approach offers a significant advance in the accuracy and efficiency of automated thyroid cancer diagnosis from ultrasound images, with strong potential for clinical application.</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

An integrated multi-task deep learning framework for thyroid ultrasound image segmentation and classification

  • 期刊论文

  • 发表日期：

    2026年1月27日

查看全文
  • Wei Xu,
  • Shuxian Zhu,
  • Zhengtian Wu,
  • Baoqing Yu

摘要

The rising incidence of thyroid cancer highlights the importance of ultrasound imaging for early detection and diagnosis. Accurate segmentation and classification of thyroid nodules in ultrasound images are essential for improving diagnostic precision and guiding clinical treatment. However, conventional deep learning methods often address these tasks separately, missing the potential benefits of tackling them together and limiting overall effectiveness. In this study, we propose a novel multi-task learning (MTL) framework with shared multi-layer parameters to perform nodule segmentation and classification simultaneously. Our architecture includes a Hierarchical Shared Network (HSN) for strong feature extraction, a Task Interaction Module (TIM) to enable dynamic information exchange between tasks, and a Dynamic Decoding Module (DDM) for enhanced segmentation. Experiments on the public TN3K and TG3K datasets demonstrate that our MTL framework achieves a classification accuracy of 93.4% and a segmentation Dice coefficient of 87.8%, outperforming several established approaches. Ablation studies further confirm the positive impact of HSN, TIM, and DDM on overall performance. These findings show that our integrated MTL approach offers a significant advance in the accuracy and efficiency of automated thyroid cancer diagnosis from ultrasound images, with strong potential for clinical application.