Scholarlink

Background <p>HER2-positive breast cancer (HER2 + BC) is clinically distinct from other subtypes, such as triple-negative or hormone receptor–positive breast cancers, due to its unique tumor microenvironment (TME) and its heterogeneous response to neoadjuvant chemotherapy (NAC). Given the critical role of the TME in treatment outcomes, we investigated whether TME features extracted from histopathological images can predict pathological complete response (pCR) and guide personalized therapy.</p> Methods <p>We retrospectively analyzed 147 HER2 + BC patients treated with NAC, including 85 from the Yale Response dataset (training cohort) and 62 from the IMPRESS HER2+ dataset (external validation cohort). Hematoxylin and eosin-stained histopathology images were segmented using VGG-16 and Xception networks to generate tissue segmentation images (TS-images). Based on the TS-images, tumor and stroma regions were segmented. Intratumoral and stromal tumor-infiltrating lymphocytes (iTILs and sTILs, respectively) were extracted from these regions and then combined to form TILs. The morphological features of these regions were quantitatively characterized using connected component analysis. Feature selection was performed by integrating morphological and clinical data via the least absolute shrinkage and selection operator. The selected features were then used to train a multilayer perceptron model, which was validated on the IMPRESS HER2+ dataset.</p> Results <p>In external validation, the model based on sTILs achieved an AUC of 0.873 for pCR prediction, with an F1 score of 0.889, PPV of 0.821, recall of 0.970, and NPV of 0.933. This performance substantially outperformed models trained on stroma (AUC = 0.779), tumor (0.732), iTILs (0.594), and TILs (0.668). Notably, the sTILs-based model maintained high performance (AUC = 0.722) even when trained with 20% of the training cohort. Univariate analyses identified morphological predictors for pCR, including the filled area of significant regions (mean) in sTILs (<i>P</i> value = 0.015).</p> Conclusion <p>Morphological TME features from histopathological images can accurately predict pCR in HER2 + BC, supporting their use in guiding NAC decision-making.</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

Morphological analysis of tumor microenvironment in HER2-positive breast cancer: predicting response to neoadjuvant chemotherapy on histopathological images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月21日

查看全文
  • Wensheng Cui,
  • Ming Fan,
  • Lihua Li

摘要

Background

HER2-positive breast cancer (HER2 + BC) is clinically distinct from other subtypes, such as triple-negative or hormone receptor–positive breast cancers, due to its unique tumor microenvironment (TME) and its heterogeneous response to neoadjuvant chemotherapy (NAC). Given the critical role of the TME in treatment outcomes, we investigated whether TME features extracted from histopathological images can predict pathological complete response (pCR) and guide personalized therapy.

Methods

We retrospectively analyzed 147 HER2 + BC patients treated with NAC, including 85 from the Yale Response dataset (training cohort) and 62 from the IMPRESS HER2+ dataset (external validation cohort). Hematoxylin and eosin-stained histopathology images were segmented using VGG-16 and Xception networks to generate tissue segmentation images (TS-images). Based on the TS-images, tumor and stroma regions were segmented. Intratumoral and stromal tumor-infiltrating lymphocytes (iTILs and sTILs, respectively) were extracted from these regions and then combined to form TILs. The morphological features of these regions were quantitatively characterized using connected component analysis. Feature selection was performed by integrating morphological and clinical data via the least absolute shrinkage and selection operator. The selected features were then used to train a multilayer perceptron model, which was validated on the IMPRESS HER2+ dataset.

Results

In external validation, the model based on sTILs achieved an AUC of 0.873 for pCR prediction, with an F1 score of 0.889, PPV of 0.821, recall of 0.970, and NPV of 0.933. This performance substantially outperformed models trained on stroma (AUC = 0.779), tumor (0.732), iTILs (0.594), and TILs (0.668). Notably, the sTILs-based model maintained high performance (AUC = 0.722) even when trained with 20% of the training cohort. Univariate analyses identified morphological predictors for pCR, including the filled area of significant regions (mean) in sTILs (P value = 0.015).

Conclusion

Morphological TME features from histopathological images can accurately predict pCR in HER2 + BC, supporting their use in guiding NAC decision-making.