Scholarlink

Objective <p>To evaluate the diagnostic value of enhanced computed tomography (CT) radiomics and deep learning in differentiating pediatric peripheral neuroblastoma (NB) from ganglioneuroblastoma (GNB).</p> Methods <p>Retrospectively enrolled children with pathologically confirmed NB and GNB between February 2014 and December 2024 were randomly sampled and divided into a training set and a validation set. Radiomic features were extracted and selected from arterial-phase and venous-phase CT images. A 3D-UNet model was constructed to train a lesion segmentation model. Radiomic features were automatically extracted and selected from CT images using the 3D-UNet model. Radiomics models and combined models (incorporating clinical indicators) were established on the basis of the optimal feature subsets for the arterial phase, venous phase, arteriovenous phase, and combined clinical data. Receiver operating characteristic (ROC) curves were plotted, and the area under the curve (AUC) was calculated and compared across phases and models.</p> Results <p>A total of 225 pediatric patients (161 NB cases and 64 GNB cases) were enrolled. Compared with the radiomic-only models (LDA and LR), the combined models (LDA and LR) incorporating radiomic and clinical features (age and NSE) demonstrated greater diagnostic performance in differentiating NB and GNB. The 3D-UNet-based deep learning model achieved optimal performance after 20 training iterations, with a best Dice coefficient of 0.818 for mixed tumor types in the validation set.The optimal diagnostic performance was achieved by the LR combined model based on venous-phase CT images’ omics features and clinical data (CE2 + C), with an AUC of 0.922 in the training set and 0.913 in the validation set. Both the LDA and LR models exhibited high diagnostic efficacy, with no statistically significant difference (<i>P</i> &gt; 0.05) between the models. There was also no significant difference (<i>P</i> &gt; 0.05) in the diagnostic performance of the combined models across phases when incorporating clinical information.</p> Conclusion <p>Enhanced CT radiomics and deep learning models have significant diagnostic value in differentiating pediatric peripheral NB from GNB.</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

The diagnostic value of enhanced CT radiomics and deep learning in differentiating pediatric peripheral neuroblastoma from ganglioneuroblastoma

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月6日

查看全文
  • Guangfeng Zhang,
  • Feng Gao,
  • Lei Fan,
  • Wenbin Guo,
  • Jianshe Zhao

摘要

Objective

To evaluate the diagnostic value of enhanced computed tomography (CT) radiomics and deep learning in differentiating pediatric peripheral neuroblastoma (NB) from ganglioneuroblastoma (GNB).

Methods

Retrospectively enrolled children with pathologically confirmed NB and GNB between February 2014 and December 2024 were randomly sampled and divided into a training set and a validation set. Radiomic features were extracted and selected from arterial-phase and venous-phase CT images. A 3D-UNet model was constructed to train a lesion segmentation model. Radiomic features were automatically extracted and selected from CT images using the 3D-UNet model. Radiomics models and combined models (incorporating clinical indicators) were established on the basis of the optimal feature subsets for the arterial phase, venous phase, arteriovenous phase, and combined clinical data. Receiver operating characteristic (ROC) curves were plotted, and the area under the curve (AUC) was calculated and compared across phases and models.

Results

A total of 225 pediatric patients (161 NB cases and 64 GNB cases) were enrolled. Compared with the radiomic-only models (LDA and LR), the combined models (LDA and LR) incorporating radiomic and clinical features (age and NSE) demonstrated greater diagnostic performance in differentiating NB and GNB. The 3D-UNet-based deep learning model achieved optimal performance after 20 training iterations, with a best Dice coefficient of 0.818 for mixed tumor types in the validation set.The optimal diagnostic performance was achieved by the LR combined model based on venous-phase CT images’ omics features and clinical data (CE2 + C), with an AUC of 0.922 in the training set and 0.913 in the validation set. Both the LDA and LR models exhibited high diagnostic efficacy, with no statistically significant difference (P > 0.05) between the models. There was also no significant difference (P > 0.05) in the diagnostic performance of the combined models across phases when incorporating clinical information.

Conclusion

Enhanced CT radiomics and deep learning models have significant diagnostic value in differentiating pediatric peripheral NB from GNB.