背景和目的

       乳腺癌是一种高度异质性的恶性肿瘤，在遗传学及基因表达方面具有显著的多样性，这导致了患者的治疗敏感性和预后存在极大差异。乳腺癌预后个体化精准评估对治疗决策至关重要：对于预后不良人群，在采取更积极的治疗方案的同时，规避伴随而来的显著副作用对改善患者预后有重要意义；对于预后好的患者，依据个体临床特征和肿瘤生物学特性进行个性化降级治疗，是避免过度治疗的关键策略。因此，个体化精准评估预后是提高乳腺癌患者治疗效益比，改善患者生存的重要途径，成为乳腺癌研究的前沿领域。

       现行的评估方法主要依赖于对术后肿瘤组织的分子和基因表达情况的测定，多项生物标志物成为判断乳腺癌预后的重要依据。然而，上述方法存在有创、耗时等局限性。近年来，研究显示通过影像学准确表征乳腺癌的生物学特征，影像生物标志物评估乳腺癌患者预后已成为研究热点。

       超声检查作为乳腺癌术前评估的重要影像学手段，能够全方位、动态且实时地获取肿瘤的形态及功能信息。临床实践中超声评估主要应用于乳腺肿瘤的良恶性诊断，针对超声影像生物标志物的预后研究方兴未艾。超声影像生物标志物可在定性和定量两个层面反映肿瘤的生物学信息，其中定性影像生物标志物简便易用，与临床病理信息的有效结合可实现对乳腺癌患者预后的综合评估，并取得较好的预测效果。同时，影像组学和人工智能的快速发展，隐藏在超声图像中的深层次、高维度信息得以被发掘，即实现定量影像生物标志物的提取，这些标志物有望为乳腺癌患者预后的精准评估提供量化依据，从而有助于制订个体化治疗方案，优化临床诊疗流程。

       基于以上乳腺癌预后个体化评估、超声影像智能分析的背景，本研究的主要目标和内容为（1）基于乳腺癌原发灶超声图像，利用深度学习自动化分析，探究超声定量影像生物标志物识别三阴性乳腺癌的效能；（2）通过对乳腺癌病灶超声图像语义特征提取，探究超声影像生物标志物对激素受体阳性乳腺癌高复发风险人群的评估价值，及其联合临床病理因素进行综合评估的效能。

方法

       第一部分：2018年4月至2019年3月共回顾性纳入经病理确诊为乳腺癌、由同一超声医师进行超声检查的145例患者（831张图像）。收集患者相应的临床及病理信息。145例患者以8:1:1的比例被分为训练集、验证集和测试集。根据免疫组化结果确定分子亚型。使用基于视觉几何（VGG）架构和自适应直方图均衡化的卷积神经网络（CNN）在训练集上进行建模以预测三阴性乳腺癌（TNBC）。在测试集中使用随机 k折交叉验证对其性能进行评估，评估指标包括受试者工作特征曲线下的面积（AUC）、准确性、灵敏性和特异性。运用t分布式随机相邻嵌入（t-SNE）分析和显著性图阐明模型的可解释性。

       第二部分：纳入2012年5月至2017年1月病理证实为T_1-3N_0-1M₀激素受体（HR）阳性、人表皮生长因子受体2（HER2）阴性乳腺癌、同时有明确Oncotype DX复发评分（RS）的连续病例523例。收集患者的临床及病理数据，并对术前超声图像进行语义特征提取。将患者按照入组时间分组，2012年5月至2015年12月为训练集，2016年1月至2017年1月为验证集。在训练集中进行了单变量和多变量回归分析，确定高复发风险（RS≥26）乳腺癌的独立预测因素，在此基础上，开发联合超声特征和临床病理信息的综合预测模型，以及单纯临床病理模型。用受试者工作特征曲线下面积（AUC）、校准曲线和决策曲线分析（DCA）在训练集中进行模型效能评估和比较，最终构建列线图，并在验证集中进行效能评估。

结果和讨论

       第一部分：145例患者中有16例（11.03%）TNBC。115例患者（平均年龄为52.04岁±11.42岁）为训练集，15例患者（平均年龄为49.40岁±10.15岁）为验证集，15名患者（平均年龄为50.93岁±10.07岁）为测试集。CNN模型在区分三阴性乳腺癌和其他三种亚型方面表现出良好的效果，其AUC为0.86（95%CI：0.64，0.95），准确性为85%，灵敏性为86%，特异性为86%，F1分数为0.74。此外，可视化分析显示该模型学习到的内部特征在不同分子亚型组之间有明显的差异。本研究建立的深度学习系统可自动提取超声影像特征，准确识别三阴性乳腺癌。

       第二部分：363名患者构成训练集，160名患者构成验证集。高复发风险（RS：26-100）率分别为14%和13.1%。多因素分析显示病灶超声下高回声晕、后方回声增强、肿瘤低水平孕酮受体表达（PR）和高Ki-67 指数为高 RS的独立风险因素（所有因素P＜0.05）。与单纯临床病理模型相比，综合模型具有更好的区分度（AUC，0.95[95%CI：0.93, 0.97]vs0.89 [95% CI：0.86, 0.92]，P=0.001），且在 DCA 中观察到更多的临床获益。基于综合模型开发了列线图，在验证集中AUC可达0.90 [95% CI：0.84, 0.94]。特别是在 T₁N₀M₀患者中，该列线图同样有效（AUC：0.91[95% CI：0.84, 0.95]）。高回声晕和后方回声增强是预测T_1-3N_0-1M₀、HR阳性、HER2阴性乳腺癌高复发风险患者的重要影像生物标志物。结合 PR 状态、Ki-67 指数和超声影像生物标记物的列线图在早期选择高复发风险患者，尤其在T₁N₀M₀患者中显示出良好的鉴别能力。

结论

       综上所述，本研究针对乳腺癌预后个体化评估这一临床问题，采用乳腺癌原发灶超声图像，识别出乳腺癌中的预后不良人群。一方面我们通过卷积神经网络，挖掘超声图像中的高维特征，利用超声定量影像生物标志物构建模型，实现术前对三阴性乳腺癌的预测；另一方面，我们对乳腺癌病灶进行了超声语义特征提取，联合临床病理信息构建综合模型，相较单纯临床模型，综合模型识别高复发风险患者准确性更高，有望为挑选高复发风险人群进行辅助化疗治疗获益提供有效评估手段。

Background and Purpose

Breast cancer is a highly heterogeneous malignant tumor with significant genetic and genotypic diversity, which leads to great variation in treatment sensitivity and prognosis of patients. Individualized and precise prognostic assessment of breast cancer is crucial for treatment decision-making: for those with poor prognosis, avoiding significant side effects while adopting more aggressive treatment decisions is important to improve the prognosis of patients; for patients with good prognosis, downgrading treatment according to individual clinical characteristics and tumor biology is a key strategy to avoid over-treatment. Therefore, accurate assessment of breast cancer prognosis can effectively increase the benefit ratio of treatment and improve patients’ survival.

Current methods of prognostic assessment mainly rely on molecular and genetic phenotyping of postoperative tumor tissues, and several biomarkers have become an important basis for determining the prognosis of breast cancer. However, these methods have limitations such as being invasive and time-consuming. Recently, studies have shown that imaging biomarkers have become a hot topic for assessing the prognosis of breast cancer patients by accurately characterizing the molecular features of breast cancer through imaging.

As an important imaging tool for preoperative evaluation of breast cancer, ultrasound is able to obtain morphological and functional information of the tumor in a comprehensive, dynamic and real-time manner. In clinical practice, ultrasound is mainly used for the qualitative diagnosis of breast masses, and the prognostic research on ultrasound is still in the ascendant. Ultrasound imaging biomarkers can reflect the biological information of tumors at both qualitative and quantitative levels, among which qualitative imaging biomarkers are simple and easy to use, and the effective combination with clinicopathological information can realize the comprehensive assessment of breast cancer patients’ prognosis. Meanwhile, with the rapid development of radiomics and artificial intelligence, the deep and high-dimensional information hidden in ultrasound images can be extracted, namely quantitative imaging biomarkers, which are expected to provide a quantitative basis for the accurate assessment of the prognosis, which will help to make individualized treatment plans and optimize clinical diagnosis and treatment processes.

Based on the above background of individualized assessment of breast cancer prognosis and intelligent analysis of ultrasound images, the main goals and contents of this study are (1) to explore the predictive efficacy of ultrasound quantitative imaging biomarkers for triple-negative breast cancer based on ultrasound images of breast cancer primary lesions using deep learning; (2) to explore the predictive efficacy of ultrasound qualitative imaging biomarkers on Hormone Receptor (HR)-positive breast cancer in the Oncotype DX high recurrence risk group, and its efficacy in combination with clinicopathologic factors for comprehensive assessment.

Methods

Part I: From April 2018 to March 2019, a total of 145 patients and 831 pictures were enrolled retrospectively. Clinical data and ultrasound images were gathered. Based on the results of immunohistochemistry (IHC), molecular subtypes were identified. subsequently, a CNN with an architecture based on VGG was employed to identify TNBC. Randomized k-fold stratified cross-validation was used to assess the model's performance. Saliency maps and an t-SNE analysis were employed to visualize the model.

Part II: In this retrospective study, consecutive T_1-3N_0-1M₀ breast cancer patients with Hormone Receptor-positive, Human Epidermal Growth Factor Receptor 2 -negative who had available Oncotype DX RS were reviewed. Patients were divided into a training cohort between May 2012 and December 2015 and a validation cohort between January 2016 to January 2017. Preoperative ultrasound scans were examined, and clinicopathologic data were gathered. The training cohort's independent predictors of high-risk breast cancer were assessed using univariate and multivariable regression analyses. A nomogram was also created and assessed using the area under the receiver operating characteristic curve (AUC), calibration curve, and decision curve analysis (DCA).

Results and Discussions

Part I: TNBC was identified in 16 of 145 (11.03%) patients. One hundred fifteen (80%) patients, 15 (10%) patients, and 15 (10%) patients formed the train, validation, and test set respectively. The deep learning system exhibits good efficacy, with an AUC of 0.86 (95% CI: 0.64, 0.95), an accuracy of 85%, a sensitivity of 86%, a specificity of 86%, and an F1-score of 0.74. In addition, the internal representation features learned by the model showed clear differentiation across molecular subtype groups. Such a deep learning system can automatically predict triple-negative breast cancer preoperatively and accurately. It may help to get to more precise and comprehensive management.

Part II: The training group consisted of 363 patients, while the validation cohort consisted of 160 individuals. The high rate of RS (RS: 26-100) was 14% and 13.1%, in that order. The following were found to be independent risk factors for high RS: enhanced posterior echo, low progesterone receptor level, high Ki-67 Index, and echogenic halo (all P<0.05). Based on the combined model, the nomogram was created. The results indicated greater clinical benefit in DCA and better discrimination when compared to the clinicopathological model (AUC, 0.95 [95% Confidence Interval (95% CI): 0.93, 0.97] versus 0.89 [95% CI: 0.86, 0.92], respectively; P = 0.001). Additionally, the validation cohort tested positive for it (AUC: 0.90 [95% CI: 0.84, 0.94]), with the T1N0M0 patients showing the greatest benefit (AUC: 0.91 [95% CI: 0.84, 0.95]). Ultrasound features served as valuable imaging biomarkers for prediction high-risk patients in breast cancer with T_1-3N_0-1M₀, Hormone Receptor-positive, Human Epidermal Growth Factor Receptor 2-negative. A nomogram incorporating PR status, Ki-67 index, and ultrasound imaging biomarkers showed a good discrimination ability in early selection of patients at high risk, especially in T₁N₀M₀ patient.

Conclusion

In summary, this study aimed at individualized assessment of breast cancer prognosis and conducted clinical trials to identify the population with poor prognosis in early-stage breast cancer by extracting semantic features and mining deep features. On the one hand, we used convolutional neural networks to deeply mine high-dimensional features in ultrasound images and established a model based solely on ultrasound quantitative imaging biomarkers to achieve prediction of triple-negative breast cancer; on the other hand, we extracted ultrasound qualitative imaging biomarkers of breast cancer lesions, and constructed a comprehensive model combined with clinicopathological information. Compared with clinicopathological model, it has higher clinical benefits for patients with high RS and is expected to provide an effective evaluation method for selecting groups with high-risk of recurrence for adjuvant chemotherapy benefits.