Scholarlink

Background <p>The accurate preoperative identification of axillary lymph node (ALN) metastasis status is important for treatment decision-making and prognostic assessment for patients with breast cancer. The aim of this study was to develop a multiscale feature fusion model (CBAM_DenseNet201_RC), which integrates attention-guided deep learning from ultrasound, radiomics, and clinical features for the preoperative prediction of ALN status in patients with breast cancer.</p> Methods <p>Retrospectively, 323 eligible breast cancer patients in our hospital between January 2018 and December 2021 were analyzed and included, and randomly divided into a training group (226 patients) and test group (97 patients) according to a 7:3 ratio. First, the initial prediction of the DenseNet201 baseline deep learning model was applied on the pre-processed completed ultrasound images of the training and test groups, and the Convolutional Block Attention Module (CBAM), which sequentially applies both spatial and channel weighting, was integrated to obtain the attention-enhanced deep learning features. Then, the radiomics features of the preprocessed ultrasound were extracted separately, and the Mann-Whitney U-test, random forest recursive feature elimination, least absolute shrinkage and selection operator three-step methods were utilized to reduce the dimensionality of the features and construct the radiomics signatures. Single/multifactor analysis was used to screen valuable clinical features. Finally, the attention-enhanced deep learning features, radiomics signatures and valuable clinical features were sequentially input into a logistic regression classifier to construct a multi-scale feature fusion model.</p> Results <p>On the test group, the AUC of DenseNet201 was 0.771 (95% CI: 0.675–0.856). the AUC of the CBAM-guided DenseNet201 model (CBAM_DenseNet201) could be further improved to 0.795 (0.692–0.878). In contrast, CBAM_DenseNet201_RC achieved the best predictive performance with an AUC as high as 0.834 (0.741–0.906), which outperformed all the sub-models, with an AUC of 0.797 (0.701–0.880) for the radiomics model and an AUC of 0.707 (0.589–0.809) for the clinical model.</p> Conclusions <p>A multiscale feature fusion model constructed by ultrasound-based attention-guided deep learning in conjunction with radiomics can effectively, noninvasively, and accurately predict ALN metastasis, enabling radiologists to formulate early diagnosis and treatment plans and potentially improving the survival rate for patients with breast cancer.</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

Ultrasound-based attention-guided deep learning combined with radiomics to predict axillary lymph node metastasis in breast cancer

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月9日

查看全文
  • Heng Zhang,
  • Tong Zhao,
  • Sai Zhang,
  • Zi-Yi Wang,
  • Nan-Nan Cao,
  • Jiang-Yi Ding,
  • Xiao-Qin Li,
  • Xiao-Li Zhou,
  • Xin-Ye Ni

摘要

Background

The accurate preoperative identification of axillary lymph node (ALN) metastasis status is important for treatment decision-making and prognostic assessment for patients with breast cancer. The aim of this study was to develop a multiscale feature fusion model (CBAM_DenseNet201_RC), which integrates attention-guided deep learning from ultrasound, radiomics, and clinical features for the preoperative prediction of ALN status in patients with breast cancer.

Methods

Retrospectively, 323 eligible breast cancer patients in our hospital between January 2018 and December 2021 were analyzed and included, and randomly divided into a training group (226 patients) and test group (97 patients) according to a 7:3 ratio. First, the initial prediction of the DenseNet201 baseline deep learning model was applied on the pre-processed completed ultrasound images of the training and test groups, and the Convolutional Block Attention Module (CBAM), which sequentially applies both spatial and channel weighting, was integrated to obtain the attention-enhanced deep learning features. Then, the radiomics features of the preprocessed ultrasound were extracted separately, and the Mann-Whitney U-test, random forest recursive feature elimination, least absolute shrinkage and selection operator three-step methods were utilized to reduce the dimensionality of the features and construct the radiomics signatures. Single/multifactor analysis was used to screen valuable clinical features. Finally, the attention-enhanced deep learning features, radiomics signatures and valuable clinical features were sequentially input into a logistic regression classifier to construct a multi-scale feature fusion model.

Results

On the test group, the AUC of DenseNet201 was 0.771 (95% CI: 0.675–0.856). the AUC of the CBAM-guided DenseNet201 model (CBAM_DenseNet201) could be further improved to 0.795 (0.692–0.878). In contrast, CBAM_DenseNet201_RC achieved the best predictive performance with an AUC as high as 0.834 (0.741–0.906), which outperformed all the sub-models, with an AUC of 0.797 (0.701–0.880) for the radiomics model and an AUC of 0.707 (0.589–0.809) for the clinical model.

Conclusions

A multiscale feature fusion model constructed by ultrasound-based attention-guided deep learning in conjunction with radiomics can effectively, noninvasively, and accurately predict ALN metastasis, enabling radiologists to formulate early diagnosis and treatment plans and potentially improving the survival rate for patients with breast cancer.