Scholarlink

<p>Accurate and efficient diagnosis of thyroid cancer through ultrasonography poses a valuable yet challenging task for practicing radiologists, given the complexity and variations present in thyroid images. Numerous automated methods have been introduced to aid radiologists in this regard. However, a significant limitation of many segmentation methods is the precise delineation of thyroid nodules from images. This challenge is primarily attributed to uneven information distribution within a region, the presence of surrounding organs and tissues with similar intensity values, variations in size and texture, and issues related to class imbalance. Current methodologies strive to tackle these challenges through diverse mechanisms; however, a void in addressing this task persists. To address these challenges, we have proposed an Improved UNet network namely Thyroid UNet (Th-UNet) designed for automatic thyroid nodule segmentation. Th-UNet incorporates a novel weighted dual path feature integration module (WDPFIM) to effectively reuse local and global information across the network. These aggregating features of different scales, considering their hierarchical importance, enhances spatial precision and localization information throughout the network. To minimize the loss of valuable information because of pooling operation, we have replaced the conventional max-pooling layer with novel augmented max pooling (AMP) layers. Similarly, to mitigate issues related to class imbalance, this study employed a modified version of a hybrid loss function. The Th-UNet uses interlinked decoder to capture both fine-scale and coarse-scale features to improve the extraction of thyroid nodules. Finally, a contrast improvement module namely (CIM) was integrated into the proposed model that selectively enhances fine details of nodule while avoiding unnecessarily amplification of non-nodule regions. The proposed model underwent evaluation on four distinct thyroid image datasets, demonstrating superior performance compared to state-of-the-art methods in thyroid nodule segmentation. Th-UNet achieved the highest precision score of 0.854, a recall of 0.884, Dice Similarity Coefficient (DSC) of 87.99%, Jaccard index of 85.21%, Positive Agreement (PA) of 95.75%, False Positive Rate (FPR) and False Negative Rate (FNR) of 0.171, 0.137 respectively on the Unified dataset of thyroid nodule images.</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

Thyroid nodule segmentation in medical images via an improved UNet model

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Mubina Zaka,
  • Syed Hamad Shirazi,
  • Assad Rasheed,
  • Atef Masmoudi

摘要

Accurate and efficient diagnosis of thyroid cancer through ultrasonography poses a valuable yet challenging task for practicing radiologists, given the complexity and variations present in thyroid images. Numerous automated methods have been introduced to aid radiologists in this regard. However, a significant limitation of many segmentation methods is the precise delineation of thyroid nodules from images. This challenge is primarily attributed to uneven information distribution within a region, the presence of surrounding organs and tissues with similar intensity values, variations in size and texture, and issues related to class imbalance. Current methodologies strive to tackle these challenges through diverse mechanisms; however, a void in addressing this task persists. To address these challenges, we have proposed an Improved UNet network namely Thyroid UNet (Th-UNet) designed for automatic thyroid nodule segmentation. Th-UNet incorporates a novel weighted dual path feature integration module (WDPFIM) to effectively reuse local and global information across the network. These aggregating features of different scales, considering their hierarchical importance, enhances spatial precision and localization information throughout the network. To minimize the loss of valuable information because of pooling operation, we have replaced the conventional max-pooling layer with novel augmented max pooling (AMP) layers. Similarly, to mitigate issues related to class imbalance, this study employed a modified version of a hybrid loss function. The Th-UNet uses interlinked decoder to capture both fine-scale and coarse-scale features to improve the extraction of thyroid nodules. Finally, a contrast improvement module namely (CIM) was integrated into the proposed model that selectively enhances fine details of nodule while avoiding unnecessarily amplification of non-nodule regions. The proposed model underwent evaluation on four distinct thyroid image datasets, demonstrating superior performance compared to state-of-the-art methods in thyroid nodule segmentation. Th-UNet achieved the highest precision score of 0.854, a recall of 0.884, Dice Similarity Coefficient (DSC) of 87.99%, Jaccard index of 85.21%, Positive Agreement (PA) of 95.75%, False Positive Rate (FPR) and False Negative Rate (FNR) of 0.171, 0.137 respectively on the Unified dataset of thyroid nodule images.