Scholarlink

Rationale and objectives <p>The development of new antibody-drug conjugates necessitates precise identification of human epidermal growth factor receptor 2 (HER2)-low breast cancers. This study explores the potential application of an automated machine learning model to develop unenhanced apparent diffusion coefficient (ADC)-based radiomics for distinguishing between HER2-zero and HER2-low/ positive breast cancers, as well as HER2-zero and HER2-low breast cancers.</p> Materials and methods <p>This study included 169 patients with invasive breast cancer (27 HER2-zero, 96 HER2-low and 46 HER2-positive). For each lesion, 1200 radiomics features and 11 clinicopathologic features were extracted. An automated machine learning pipeline was constructed using PyCaret for two binary classifications: HER2-zero versus HER2-low and HER2-zero versus HER2-low/ positive breast cancers. The performance of 13 machine learning models was assessed based on the training cohort and was compared using the area under the curve (AUC) to obtain the optimal model. This optimal model was subsequently validated using an internal validation cohort. Statistical analyses included Students’ t-tests, Mann-Whitney U tests or Chi-Square tests. The performance of the best model was evaluated using accuracy, AUC, sensitivity, positive predictive value and F1 score.</p> Results <p>In distinguishing HER2-zero from HER2-low breast cancers, the Gradient Boosting Classifier exhibited superior classification performance with an AUC of 0.75 in the validation cohort. The Random Forest Classifier (RF) exhibited the best performance for the clinical-radiomics model, achieving an AUC of 0.79. For differentiating HER2-zero from HER2-low/ positive breast cancers, RF was optimal for both radiomics and radiomics-clinicopathologic models, each achieving an AUC of 0.82 in the validation cohort.</p> Conclusion <p>The application of automated machine learning with ADC-based radiomics for the identification of HER2-low breast cancers is both feasible and advantageous.</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

Automated machine learning-based unenhanced radiomics for identification of HER2-low breast cancers using diffusion-weighted imaging

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月5日

查看全文
  • Xue Li,
  • Chunmei Li,
  • Lei Jiang,
  • Dandan Zheng,
  • Song Tian,
  • Min Chen

摘要

Rationale and objectives

The development of new antibody-drug conjugates necessitates precise identification of human epidermal growth factor receptor 2 (HER2)-low breast cancers. This study explores the potential application of an automated machine learning model to develop unenhanced apparent diffusion coefficient (ADC)-based radiomics for distinguishing between HER2-zero and HER2-low/ positive breast cancers, as well as HER2-zero and HER2-low breast cancers.

Materials and methods

This study included 169 patients with invasive breast cancer (27 HER2-zero, 96 HER2-low and 46 HER2-positive). For each lesion, 1200 radiomics features and 11 clinicopathologic features were extracted. An automated machine learning pipeline was constructed using PyCaret for two binary classifications: HER2-zero versus HER2-low and HER2-zero versus HER2-low/ positive breast cancers. The performance of 13 machine learning models was assessed based on the training cohort and was compared using the area under the curve (AUC) to obtain the optimal model. This optimal model was subsequently validated using an internal validation cohort. Statistical analyses included Students’ t-tests, Mann-Whitney U tests or Chi-Square tests. The performance of the best model was evaluated using accuracy, AUC, sensitivity, positive predictive value and F1 score.

Results

In distinguishing HER2-zero from HER2-low breast cancers, the Gradient Boosting Classifier exhibited superior classification performance with an AUC of 0.75 in the validation cohort. The Random Forest Classifier (RF) exhibited the best performance for the clinical-radiomics model, achieving an AUC of 0.79. For differentiating HER2-zero from HER2-low/ positive breast cancers, RF was optimal for both radiomics and radiomics-clinicopathologic models, each achieving an AUC of 0.82 in the validation cohort.

Conclusion

The application of automated machine learning with ADC-based radiomics for the identification of HER2-low breast cancers is both feasible and advantageous.