Scholarlink

<p>Ovarian cancer refers to a malignant tumor that grows in the ovary, which has the highest mortality rate of gynecological cancers. Positron emission tomography and computer tomography (PET/CT) imaging is widely used for the localization and characterization of ovarian tumors, but its analysis is susceptible to the subjectivity of clinicians. A deep learning-based PET/CT image diagnosis approach was investigated to achieve the segmentation and classification of ovarian tumors. We proposed several traditional convolutional neural networks (CNNs) for automatic ovarian tumors segmentation and classification. For segmentation, we design a hybrid network that integrates U-Net-MobileNetv3 with KiteNet to jointly capture global structure and lesion edge details, further enhanced by lesion-aware CarveMix augmentation and Dice-CE loss. For classification, we adopt ConvNeXt as the backbone and improve its robustness via Mixup data augmentation. Our method achieves a Dice coefficient of 0.826 and an accuracy of 0.912, outperforming all baseline models including U-Net, Deeplabv3, DenseNet, and Swin-Transformer. 1228 PET/CT images were employed to train and evaluate the CNN approach. Our segmentation model obtained a Dice of 0.826 compared to 0.773 obtained by U-Net-MobileNetv3, 0.444 by KiteNet, 0.634 by FCN, 0.744 by Deeplabv3, 0.667 by U-Net, and 0.747 by U-Net-VGG. Our classification model achieved the ACC of 0.912 compared to 0.907 achieved by ConvNeXt, 0.898 by DenseNet, 0.895 by EfficientNet and 0.848 by Swin-Transformer. In this study, we developed a novel deep learning framework for the simultaneous segmentation and classification of ovarian tumors in PET/CT imaging. Our key findings are: (1) A hybrid segmentation architecture that integrates U-Net-MobileNetv3 with KiteNet, enhanced by lesion-aware CarveMix augmentation and Dice-CE loss, achieved a Dice coefficient of 0.826, significantly outperforming standard models such as U-Net (0.667) and Deeplabv3 (0.744); (2) A classification pipeline based on ConvNeXt with Mixup data augmentation attained an accuracy of 0.912, surpassing DenseNet (0.898) and Swin-Transformer (0.848). These results demonstrate that our method can effectively support clinicians in both precise tumor delineation and reliable benign/malignant differentiation, offering a promising tool for intelligent ovarian cancer diagnosis.</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

Intelligent diagnosis of ovarian cancer in PET/CT imaging based on KiteNet-MobileNetv3 fusion and CarveMix augmentation

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年4月14日

查看全文
  • Junwei Li,
  • Kai Hu,
  • Xueru Fan,
  • Sisheng Wang,
  • Tao He,
  • Caiyun Xu,
  • Qiwen Cai,
  • Jinyan Chen,
  • Yating Ling,
  • Jianmin Yang,
  • Xiaohui Fan,
  • Dexing Kong,
  • Lixia Zhang,
  • Lu Li

摘要

Ovarian cancer refers to a malignant tumor that grows in the ovary, which has the highest mortality rate of gynecological cancers. Positron emission tomography and computer tomography (PET/CT) imaging is widely used for the localization and characterization of ovarian tumors, but its analysis is susceptible to the subjectivity of clinicians. A deep learning-based PET/CT image diagnosis approach was investigated to achieve the segmentation and classification of ovarian tumors. We proposed several traditional convolutional neural networks (CNNs) for automatic ovarian tumors segmentation and classification. For segmentation, we design a hybrid network that integrates U-Net-MobileNetv3 with KiteNet to jointly capture global structure and lesion edge details, further enhanced by lesion-aware CarveMix augmentation and Dice-CE loss. For classification, we adopt ConvNeXt as the backbone and improve its robustness via Mixup data augmentation. Our method achieves a Dice coefficient of 0.826 and an accuracy of 0.912, outperforming all baseline models including U-Net, Deeplabv3, DenseNet, and Swin-Transformer. 1228 PET/CT images were employed to train and evaluate the CNN approach. Our segmentation model obtained a Dice of 0.826 compared to 0.773 obtained by U-Net-MobileNetv3, 0.444 by KiteNet, 0.634 by FCN, 0.744 by Deeplabv3, 0.667 by U-Net, and 0.747 by U-Net-VGG. Our classification model achieved the ACC of 0.912 compared to 0.907 achieved by ConvNeXt, 0.898 by DenseNet, 0.895 by EfficientNet and 0.848 by Swin-Transformer. In this study, we developed a novel deep learning framework for the simultaneous segmentation and classification of ovarian tumors in PET/CT imaging. Our key findings are: (1) A hybrid segmentation architecture that integrates U-Net-MobileNetv3 with KiteNet, enhanced by lesion-aware CarveMix augmentation and Dice-CE loss, achieved a Dice coefficient of 0.826, significantly outperforming standard models such as U-Net (0.667) and Deeplabv3 (0.744); (2) A classification pipeline based on ConvNeXt with Mixup data augmentation attained an accuracy of 0.912, surpassing DenseNet (0.898) and Swin-Transformer (0.848). These results demonstrate that our method can effectively support clinicians in both precise tumor delineation and reliable benign/malignant differentiation, offering a promising tool for intelligent ovarian cancer diagnosis.