Deep Learning Predicts HER2 Status in Breast Cancer Using Ultrasound and MRI

Deep Learning Predicts HER2 Status in Breast Cancer Using Ultrasound and MRI

August 25, 2025 Author Summary 0

Understanding HER2 and Its Role in Breast Cancer

Human epidermal growth factor receptor type 2 (HER2) is one of the most important biomarkers in breast cancer. Patients with HER2-positive invasive breast cancer are more likely to benefit from neoadjuvant chemotherapy (NAC) compared to those with HER2-negative disease.

HER2 status is usually determined before surgery through immunohistochemical (IHC) analysis of core needle biopsy samples. HER2-positive breast cancer is defined as an IHC score of 3+, or an IHC score of 2+ confirmed by fluorescence in situ hybridization (FISH). HER2-negative status is defined as IHC scores of 0 or 1+, or IHC 2+ with a negative FISH result.

However, this standard approach has limitations. HER2-positive breast cancers are often highly heterogeneous, and a core needle biopsy may not capture the full HER2 profile of the tumor. Moreover, HER2 expression can vary within and between lesions over time. This variability can affect treatment planning, creating a need for more comprehensive, non-invasive assessment methods.

Ultrasonography (US) and magnetic resonance imaging (MRI) are the most commonly used imaging techniques in breast cancer diagnosis. Previous studies have shown that imaging-derived features can predict HER2 status, but conventional radiomics approaches often require manual feature extraction, which can be operator-dependent. Deep learning (DL) can automatically extract features from medical images, potentially improving prediction accuracy.

The Aim of the Study

Researchers from Daping Hospital, Army Medical University, and Jinfeng Laboratory sought to develop deep learning models using ultrasound and MRI to predict HER2 status in invasive breast cancer before surgery.

The study included 197 women with pathologically confirmed invasive breast cancer who underwent both US and MRI between January 2021 and July 2024.

Methods in Brief

Inclusion criteria:

  • Age over 18

  • Pathologically confirmed invasive breast cancer with a clearly documented HER2 grade

  • Both US and MRI within two weeks

Exclusion criteria:

  • Neoadjuvant chemotherapy or biopsy before imaging

  • Poor image quality

  • Lesion size under 5 mm

  • Lack of enhancement in MRI contrast sequences

Imaging protocols:

  • US: Greyscale and color Doppler images obtained in two orthogonal planes, BI-RADS features extracted by two radiologists.

  • MRI: Dynamic contrast-enhanced T1-weighted and T2-weighted fat-suppressed sequences, with tumor characteristics recorded according to BI-RADS MRI lexicon.

Deep learning models:

  • DL-US: ConvNeXt V2 model trained on tumor regions from US images.

  • DL-MRI: 3D ResNet18 model trained on segmented tumor volumes from MRI.

  • DL-US&MRI: Logistic regression combining predictions from the US and MRI models.

The dataset was split into training (149 patients) and test (48 patients) cohorts using stratified sampling.

Key Results

In the test cohort:

  • DL-US achieved an AUC of 0.842, sensitivity of 89.5%, and specificity of 79.3%.

  • DL-MRI achieved an AUC of 0.800, sensitivity of 78.9%, and specificity of 79.3%.

  • DL-US&MRI achieved the highest AUC of 0.898, with perfect specificity (100%) but lower sensitivity (63.2%).

Although the combined model outperformed both individual models, the differences were not statistically significant due to the limited sample size.

Interpretation and Practical Applications

The high sensitivity of the DL-US model suggests its potential as a screening tool. The DL-US&MRI model’s perfect specificity could make it valuable for confirmatory diagnosis, reducing false positives and unnecessary interventions.

The authors note that this is the first published model to combine US and MRI for HER2 prediction in invasive breast cancer. This approach leverages the strengths of each imaging modality:

  • US: Good for visualizing superficial features and vascularity.

  • MRI: Better for capturing tumor spatial heterogeneity.

Relation to Previous Research

Earlier studies have used either MRI or US separately to predict HER2 status, often with conventional radiomics methods. Reported AUCs for MRI-based models have ranged from 0.68 to 0.84, while US-based models have shown AUCs from 0.725 to 0.844.

The 3D deep learning MRI model in this study achieved an AUC of 0.800 without including clinical variables, bridging the gap between purely radiomic and combined clinical-radiomic approaches.

Limitations and Future Work

The study has several limitations:

  • Single-center dataset limits generalizability.

  • Small sample size prevented subgroup analysis of HER2-zero, HER2-low, and HER2-high cases.

  • No integration of clinical or genomic data into the models.

  • No external validation performed.

Future work should focus on multicenter studies with larger, more diverse cohorts, incorporating clinical markers and genomic data to further improve predictive performance. Accurate detection of HER2-low status is particularly important, as recent evidence suggests these patients may benefit from targeted therapy such as trastuzumab deruxtecan.

Conclusion

Deep learning models based on ultrasound and MRI show strong potential for non-invasive preoperative prediction of HER2 status in invasive breast cancer. Combining both imaging modalities improved predictive accuracy and could help guide treatment decisions, especially in cases where biopsy results are inconclusive or difficult to obtain.

 

The translation of the preceding English text in Chinese:

 

了解HER2及其在乳腺癌中的作用

人类表皮生长因子受体2型（HER2）是乳腺癌中最重要的生物标志物之一。与HER2阴性患者相比，HER2阳性的浸润性乳腺癌患者更有可能从新辅助化疗（NAC）中获益。

HER2状态通常在手术前通过对粗针穿刺活检标本进行免疫组织化学（IHC）分析来确定。HER2阳性乳腺癌被定义为IHC评分为3+，或IHC评分为2+并经荧光原位杂交（FISH）检测确认阳性。HER2阴性状态被定义为IHC评分为0或1+，或IHC评分为2+且FISH结果为阴性。

然而，这一标准方法存在局限性。HER2阳性乳腺癌常常具有高度异质性，粗针穿刺活检可能无法全面反映肿瘤的HER2特征。此外，HER2表达在病灶内和病灶之间会随时间变化。这种变异性会影响治疗方案的制定，因此需要更全面、无创的评估方法。

超声（US）和磁共振成像（MRI）是乳腺癌诊断中最常用的影像技术。既往研究表明，影像特征可以预测HER2状态，但传统放射组学方法通常需要人工特征提取，容易受操作人员影响。深度学习（DL）可以自动从医学影像中提取特征，有望提高预测准确性。

研究目的

来自陆军军医大学大坪医院和金凤实验室的研究人员旨在开发基于超声和MRI的深度学习模型，以在手术前预测浸润性乳腺癌的HER2状态。

本研究纳入了2021年1月至2024年7月期间接受超声和MRI检查的197例经病理确诊的浸润性乳腺癌女性患者。

方法简述

纳入标准：

  • 年龄大于18岁

  • 病理确诊的浸润性乳腺癌并有明确的HER2分级

  • 在两周内完成超声和MRI检查

排除标准：

  • 影像检查前接受过新辅助化疗或活检

  • 图像质量差

  • 病灶直径小于5毫米

  • MRI对比增强序列中无强化表现

影像学方案：

  • 超声：获取灰阶和彩色多普勒图像，分别在两个正交平面成像，由两位放射科医师提取BI-RADS特征。

  • MRI：获取动态增强T1加权和脂肪抑制T2加权序列，按BI-RADS MRI词汇记录肿瘤特征。

深度学习模型：

  • DL-US：基于超声肿瘤区域训练的ConvNeXt V2模型。

  • DL-MRI：基于MRI分割肿瘤体积训练的3D ResNet18模型。

  • DL-US&MRI：使用逻辑回归结合超声和MRI模型的预测结果。

数据集按分层抽样分为训练集（149例）和测试集（48例）。

主要结果

在测试集中：

  • DL-US模型的AUC为0.842，敏感性为89.5%，特异性为79.3%。

  • DL-MRI模型的AUC为0.800，敏感性为78.9%，特异性为79.3%。

  • DL-US&MRI模型的AUC最高，为0.898，特异性达到100%，但敏感性较低（63.2%）。

尽管组合模型优于单一模型，但由于样本量有限，差异无统计学意义。

解读与临床应用

DL-US模型的高敏感性提示其可作为筛查工具；DL-US&MRI模型的完美特异性则显示其在确诊中有价值，可减少假阳性和不必要的干预。

作者指出，这是首个将超声与MRI结合用于预测浸润性乳腺癌HER2状态的已发表模型。该方法利用了两种影像技术的优势：

  • 超声：适于观察浅表结构和血供

  • MRI：更好地捕捉肿瘤空间异质性

与既往研究的关系

以往研究多单独使用MRI或超声预测HER2状态，且常采用传统放射组学方法。MRI模型的AUC报道范围为0.68至0.84，超声模型的AUC为0.725至0.844。本研究的3D深度学习MRI模型在未引入临床变量的情况下取得了0.800的AUC，弥合了纯放射组学与临床-放射组学结合方法之间的差距。

局限性与未来方向

本研究局限性包括：

  • 单中心数据限制了推广性

  • 样本量较小，无法对HER2-0、HER2低表达和HER2高表达进行亚组分析

  • 模型未纳入临床或基因组数据

  • 未进行外部验证

未来研究应开展多中心、大样本研究，并整合临床及基因组数据，以进一步提高预测性能。准确检测HER2低表达尤为重要，因为最新证据表明此类患者可能从德鲁司他单抗等靶向药物中获益。

结论

基于超声和MRI的深度学习模型在无创术前预测浸润性乳腺癌HER2状态方面显示出良好潜力。结合两种影像模态可提高预测准确性，并有助于在活检结果不明确或难以获取时指导治疗决策。

Reference:

Yuhong Fan, Kaixiang Sun, Yao Xiao, Peng Zhong, Yun Meng, Yang Yang, Zhenwei Du, Jingqin Fang

Deep learning predicts HER2 status in invasive breast cancer from multimodal ultrasound and MRI.

Biomol Biomed [Internet]. 2025 May 16 [cited 2025 Aug. 14];

Available from: https://www.bjbms.org/ojs/index.php/bjbms/article/view/12475

Additional information:

We invite submissions for our upcoming thematic issues, including:

More news: Blog

Editor: Merima Hadžić

Be the first to comment

Leave a Reply