Scholarlink

<p>Thyroid cancer diagnosis from ultrasound images remains a challenging task due to the subjective nature of image interpretation and the variability among observers, which often leads to diagnostic uncertainty and unnecessary procedures. To address these issues, this study introduces CGR-Net, a robust deep learning framework designed to enhance the accuracy and reliability of thyroid nodule classification. The proposed model integrates Conditional Generative Adversarial Networks (CGAN) for class-specific data augmentation, Gaussian noise injection (GNI) for regularization, and a fine-tuned ResNet50 network for classification. Using a publicly available dataset of 3115 ultrasound images, including 1905 benign and 1210 malignant nodules, CGAN was employed to generate synthetic malignant samples to balance the dataset, while Gaussian noise with a standard deviation of 0.05 was added to simulate imaging variability. The performance of CGR-Net was evaluated against state-of-the-art models such as Swin Transformer, EfficientNet, and Vision Transformer on unaugmented test data. The proposed model achieved an accuracy of 94.66%, sensitivity of 92.59%, and specificity of 96.06%, outperforming the baseline ResNet50 by 18.16%, Swin Transformer by 15.81%, EfficientNet by 2.78%, and Vision Transformer by 4.49%. These results demonstrate that CGR-Net effectively mitigates class imbalance and overfitting, achieving high diagnostic accuracy on an independent test set.</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

Cgr-net: a robust deep learning framework for thyroid cancer classification using conditional GAN and Gaussian-regularized ResNet50

  • 期刊论文

  • 发表日期：

    2026年5月11日

查看全文
  • Jihad A. Qadir

摘要

Thyroid cancer diagnosis from ultrasound images remains a challenging task due to the subjective nature of image interpretation and the variability among observers, which often leads to diagnostic uncertainty and unnecessary procedures. To address these issues, this study introduces CGR-Net, a robust deep learning framework designed to enhance the accuracy and reliability of thyroid nodule classification. The proposed model integrates Conditional Generative Adversarial Networks (CGAN) for class-specific data augmentation, Gaussian noise injection (GNI) for regularization, and a fine-tuned ResNet50 network for classification. Using a publicly available dataset of 3115 ultrasound images, including 1905 benign and 1210 malignant nodules, CGAN was employed to generate synthetic malignant samples to balance the dataset, while Gaussian noise with a standard deviation of 0.05 was added to simulate imaging variability. The performance of CGR-Net was evaluated against state-of-the-art models such as Swin Transformer, EfficientNet, and Vision Transformer on unaugmented test data. The proposed model achieved an accuracy of 94.66%, sensitivity of 92.59%, and specificity of 96.06%, outperforming the baseline ResNet50 by 18.16%, Swin Transformer by 15.81%, EfficientNet by 2.78%, and Vision Transformer by 4.49%. These results demonstrate that CGR-Net effectively mitigates class imbalance and overfitting, achieving high diagnostic accuracy on an independent test set.