中文摘要

第一部分 超声视频全病灶感知网络技术在乳腺病灶良恶性诊断的应用研究

背景

基于乳腺超声静态图像开发的人工智能（AI）模型目前尚无法在实际临床工作中广泛应用，其中重要的原因是单帧静态图像无法显示病变全貌、且关键帧获取操作者依赖性高。高频超声具有高分辨力、实时动态、高效的技术优势，我们探索采用基于全病变感知网络（WAUVE）的深度学习模型分析超声动态扫查病灶的视频，以提高超声人工智能诊断乳腺癌的准确性。

资料与方法

回顾性收集2020年5月至2022年8月在两家医院的2771名患者的2912个乳腺病灶的超声动态扫查视频用于训练WAUVE模型。将WAUVE的诊断性能与静态2D-ResNet50模型和动态TimeSformer模型进行了比较，并在内部验证集上进行了验证。随后，2022年12月至2023年4月在另外两家医院前瞻性收集190例患者的190个乳腺病变超声动态扫查视频作为独立外部验证集。四名经验丰富的超声科医生对外部验证集的病例视频进行了独立盲法诊断。比较WAUVE与四位超声科医生的诊断效能进行比较，并评估该模型对超声医生诊断的辅助价值。

结果

WAUVE表现出优于2D-ResNet50模型的性能，而与TimeSformer模型相似。在外部验证集中，WAUVE的AUC达到了0.8998（95% CI=0.8529-0.9439），在敏感性（97.39% vs. 98.48%, p=0.36）、特异性（49.33% vs. 50.00%, p=0.92）和准确性（78.42% vs. 79.34%, p=0.60）方面与四名经验丰富的超声科医生的表现无明显差异。在WAUVE模型的帮助下，四名超声科医生的平均特异性提高了6.67%，并且诊断一致性提高（从0.807提高到0.838）。

结论

基于自由手持超声扫描视频的WAUVE深度学习模型在乳腺癌诊断中表现出优异的性能，取得的结果与经验丰富的超声科医生相似，有望实现临床应用。

 

第二部分 乳腺癌原发灶超声影像组学预测腋窝淋巴结转移肿瘤负荷

背景

基于原发性乳腺病灶超声影像组学分析，建立T1/T2期浸润性乳腺癌患者腋窝淋巴结（ALN）状态的预测模型。

资料与方法

2016年8月至2018年11月期间，共纳入343例经组织学证实的恶性乳腺肿瘤患者，按7:3比例随机分为训练集和验证集。ALN肿瘤负荷定义为低负荷（<3个转移性ALN）或高负荷（≥3个转移性ALN）。使用"PyRadiomics"包提取影像组学特征，通过最小绝对收缩与选择算子逻辑回归构建影像组学评分。基于多因素逻辑回归结果，整合乳腺癌超声影像组学评分、患者年龄及病灶大小构建列线图模型。

结果

在训练集和验证集中，分别有29.1%（69/237）和32.08%（34/106）患者病理诊断为≥3个转移性ALN。影像组学评分包含16个超声特征，联合患者年龄和病灶大小构建模型。模型在训练集的AUC为0.846（95%CI 0.790-0.902），验证集为0.733（95%CI 0.613-0.852）。校准曲线显示预测值与观察值具有良好一致性。

结论

综合乳腺癌原发灶影像组学特征、患者年龄、病灶大小构建的列线图模型有助于术前预测浸润性乳腺癌患者ALN转移的肿瘤负荷。

 

 

第三部分 早期乳腺癌前哨淋巴结超声的模态自适应网络预测腋窝淋巴结转移肿瘤负荷

背景

cT1-2N0 期乳腺癌患者的ALN肿瘤负荷对于确定治疗方案、判断预后有重要价值。通过经皮超声造影技术定位显示前哨淋巴结（SLN）后，针对SLN的超声图像，开发并验证深度学习模型直接评估ALN肿瘤负荷。

资料与方法

本研究前瞻性纳入 2020 年 4 月至 2021 年 7 月期间在北京协和医院，以及 2022 年 4 月至 2022 年 7 月期间在四川省肿瘤医院接受腋窝CEUS检查的 cT1-T2N0 期乳腺癌女性患者。我们尝试设计一种以灰阶或彩色多普勒（任意一种）超声图像结合临床先验知识（SLN长轴及短轴长度，SLN长短轴比值，淋巴结皮质厚度），联合临床病理信息（年龄、激素受体状态、Her-2受体，Ki67状态）作为输入，并输出ALN转移状态是肿瘤高负荷（转移淋巴结≥3个）概率的深度学习模型。

结果

本研究在两个中心共纳入 374 例患者的 595 枚SLN，其中118例（31.6%）有SLN转移，肿瘤高负荷患者35例（9.4%）。对比不同基础框架的预测表现后，我们选用了IBN-ResNet作为模型框架，得到了 “模态自适应网络”（MAN）。我们将临床病理信息纳入患者数据网络，整合特征融合模块后的联合模型（MAN+C）判断ALN转移淋巴结≥3个，训练集、验证集、独立测试集和外部测试集上的AUC分别为 0.91（95% CI: 0.899-0.943）、0.98（95% CI: 0.950-1）、0.89（95% CI: 0.857-0.923）和 0.84（95% CI: 0.811-0.869）。

结论

MAN+C 提供了一种针对淋巴结的直接评估方法，可用于 cT1-2N0 期乳腺癌患者ALN肿瘤负荷的术前精准评估。

 

第四部分 联合临床病理参数与超声影像组学预测乳腺癌70基因检测（MammaPrint）的风险

背景

70基因特征检测（MammaPrint）可评估激素受体阳性/人表皮生长因子受体2阴性（HR+/HER2-）早期乳腺癌患者的预后及化疗潜在获益。然而该检测的高昂费用限制了其临床应用。

资料与方法

本研究回顾性纳入178例接受70基因检测且具备合格超声图像的HR+/HER2-早期乳腺癌女性患者。由两位超声科医师独立勾画感兴趣区域（ROIs）后，采用"Pyradiomics"软件从超声图像中提取影像组学特征。通过ICC评估观察者间及观察者内一致性。使用LASSO回归筛选特征，并通过多因素logistic回归构建结合临床病理参数和影像组学评分的列线图模型预测70基因风险分类。

结果

根据70基因检测结果，82例（46.1%）患者被归类为高风险，96例（53.9%）为低风险。感兴趣区域（ROI）勾画的影像组学特征提取具有良好重复性，观察者间一致性（ICC=0.815±0.015）和观察者内一致性（ICC=0.836±0.022）均超过0.75的强可靠性阈值。多因素分析显示影像组学评分与Ki67指数与70基因风险具有显著相关性。联合模型在训练集和验证集分别获得0.899（95%CI:0.8451-0.9523）和0.846（95%CI:0.7362-0.9561）的AUC值，显著优于单纯临床模型[0.845(0.7768-0.9136)/0.806(0.686-0.9265)]和单纯影像组学模型[0.8(0.7222-0.8783)/0.662(0.5146-0.8102)]。校正曲线、成本效益曲线及临床影响曲线均证实联合模型具有最佳精准度、可靠性及成本效益。

结论

基于乳腺癌超声影像组学特征联合临床病理参数的列线图模型可有效预测70基因风险分层。有望为无法进行70基因检测的患者提供替代评估方案。

 

ABSTRACT

Whole-lesion-aware network based on freehand ultrasound video for breast cancer assessment: a prospective multicenter study

Background

The clinical application of artificial intelligence (AI) models based on breast ultrasound static images has been hindered in real-world workflows due to operator-dependence of key frame acquisition and incomplete view of breast lesions on static images. To better exploit the real-time advantages of ultrasound and more conducive to clinical application, we proposed a whole-lesion-aware network based on free-hand ultrasound video (WAUVE) scanning in an arbitrary direction for predicting overall breast cancer risk score.

Methods and Materials

The WAUVE was developed using 2912 videos (2912 lesions) of 2771 patients retrospectively collected from May 2020 to August 2022 in two hospitals. We compared the diagnostic performance of WAUVE with static 2D-ResNet50 and dynamic TimeSformer models in the internal validation set. Subsequently, a dataset comprising 190 videos (190 lesions) from 175 patients prospectively collected from December 2022 to April 2023 in two other hospitals, was used as an independent external validation set. A reader study was conducted by four experienced radiologists on the external validation set. We compared the diagnostic performance of WAUVE with the four experienced radiologists and evaluated the auxiliary value of model for radiologists.

Results

The WAUVE demonstrated superior performance compared to the 2D-ResNet50 model, while similar to the TimeSformer model. In the external validation set, WAUVE achieved an area under the receiver operating characteristic curve (AUC) of 0.8998 (95% CI=0.8529-0.9439), and showed a comparable diagnostic performance to that of four experienced radiologists in terms of sensitivity (97.39% vs. 98.48%, p=0.36), specificity (49.33% vs. 50.00%, p=0.92), and accuracy (78.42% vs.79.34%, p=0.60). With the WAUVE model assistance, the average specificity of four experienced radiologists was improved by 6.67%, and higher consistency was achieved (from 0.807 to 0.838).

Conclusions

The WAUVE based on non-standardized ultrasound scanning demonstrated excellent performance in breast cancer assessment which yielded outcomes similar to those of experienced radiologists, indicating the clinical application of the WAUVE model promising.

 

 

 

 

Nomogram based on radiomics analysis of primary breast cancer ultrasound images: prediction of axillary lymph node tumor burden in patients

Background

To establish a prediction model for evaluating the axillary lymph node (ALN) status of patients with T1/T2 invasive breast cancer based on radiomics analysis of US images of primary breast lesions.

Methods and Materials

Between August 2016 and November 2018, a total of 343 patients with histologically proven malignant breast tumors were included in this study and randomly assigned to the training and validation groups at a ratio of 7:3. ALN tumor burden was defined as low (< 3 metastatic ALNs) or high (≥ 3 metastatic ALNs). Radiomics features were obtained using the PyRadiomics package, and the radiomics score was established by least absolute shrinkage and selection operator regression. A nomogram combining the breast cancer US radiomics score with patient age and lesion size was generated based on the multivariate logistic regression results.

Results

In the training and validation cohorts, 29.1% (69/237) and 32.08% (34/106) of patients were pathologically diagnosed with ≥3 metastatic ALNs, respectively. The radiomics score consisted of 16 US features, and patient age and lesion diameter identified by US were included to construct the model. The AUC of the model was 0.846 (95% CI, 0.790-0.902) for the training cohort and 0.733 (95% CI, 0.613-0.852) for the validation cohort. The calibration curves showed good agreement between the predictions and observations.

Conclusions

Our novel nomogram demonstrates high accuracy in predicting ALN tumor burden in breast cancer patients. We also suggest further development of PyRadiomics to improve US radiomics.

 

Preoperative Assessment of Axillary Lymph Node Tumor Burden in cT1-2N0 Breast Cancer Patients with a Modality-adaptive Network Based on Sentinel Lymph Node Ultrasound Images

Background

To determine a direct method for diagnosing axillary lymph node (ALN) tumor burden preoperatively in cT1-2N0 breast cancer patients, we developed and validated a deep learning model based on US images of sentinel lymph nodes (SLNs) detected by axillary contrast-enhanced ultrasound (CEUS).

Methods and Materials

Women with cT1-T2N0 breast cancer who received axillary CEUS were enrolled prospectively from Peking Union Medical College Hospital between April 2020 and July 2021 and from Sichuan Cancer Hospital between April 2022 and July 2022. We attempted to design a deep learning model that takes grayscale or color Doppler ultrasound images(either type) combined with clinical prior knowledge (such as SLN long-axis length, short-axis length, long-to-short axis ratio, and lymph node cortical thickness) and clinicopathological information (including age, hormone receptor status, HER-2 status, and Ki67 status) as inputs, and outputs the probability of axillary lymph node metastasis being high tumor burden (≥3 metastatic lymph nodes).

Results

A total of 374 patients with 595 sentinel lymph nodes were included from the two centers, among which 118 patients (31.6%) had SLN metastasis, and 35 patients (9.4%) had high tumor burden. After comparing the predictive performance of different baseline frameworks, we selected IBN-ResNet as the model framework and developed the "Modality Adaptive Network" (MAN). By incorporating clinicopathological information into the patient data network and integrating a feature fusion module, the combined model (MAN+C) demonstrated excellent performance in identifying axillary lymph node metastasis (≥3 metastatic lymph nodes). The AUC values for the training set, validation set, independent test set, and external test set were 0.91 (95% CI: 0.899-0.943), 0.98 (95% CI: 0.950-1), 0.89 (95% CI: 0.857-0.923), and 0.84 (95% CI: 0.811-0.869), respectively.

Conclusions

MAN+C provided a direct and efficient method for accurate preoperative assessment of ALN tumor burden in cT1-2N0 breast cancer patients.

 

 

 

 

Prediction of the 70-gene signature (MammaPrint) categorized risk by combining radiomics and clinicopathological parameters among Chinese breast cancer

Background

The 70-gene signature (70-GS) assay (MammaPrint) has been endorsed by prominent guidelines for assessing the prognosis and the potential benefits of chemotherapy in individuals with early-stage breast cancer characterized by hormone receptor positivity and human epidermal growth factor receptor 2 negativity (HR+/HER2-). However, the high cost of the 70-gene test has impeded its widespread application, particularly in developing countries.

Methods and Materials

We retrospectively included a consecutive cohort of 178 female patients with HR+/Her2- early breast cancer who received 70-GS test and had eligible ultrasound images available. Two radiologists drew the regions of internet (ROIs) independently. The “Pyradiomics” was used to extract radiomics features from ultrasound images. Intraclass correlation coefficient (ICC) was used to measure the inter-observer agreement and intra-observer agreement of the two radiologists in ROI delineation. Lasso regression was used to extract radiomics features. Multivariate logistic regression was performed to establish the nomogram models combining clinicopathological information and radiomics score to predict the categorized risk of 70-GS.

Results

According to the 70-gene signature (70-GS) classification, 82 patients were categorized as high risk and 96 as low risk. The inter-observer agreement and intra-observer agreement of the radiomics feature extraction in ROI delineation was substantial, with an ICC of 0.815±0.015, and 0.836±0.022, both exceeding the threshold of 0.75 for strong reliability. The radiomics score and Ki67 showed the strong associations with the 70-GS risk from the multivariate analysis forest plot. Compared to models using only clinicopathological characteristics or only the radiomics score, the combined (radio-clinical) model demonstrated the best performance, achieving the area under curve (AUC) of receiver-operating curve (ROC) of 0.899 (95% confidence interval (CI): 0.8451-0.9523) in the training cohort and 0.846 (95% CI: 0.7362-0.9561) in the validation cohort. While that of clinical model and radiomics model are 0.845 (95%CI: 0.7768-0.9136) and 0.8 (95%CI: 0.7222-0.8783) in the training cohort, 0.806 (95%CI: 0.686-0.9265) and 0.662 (95%CI: 0.5146-0.8102) in the validation cohort. Based on the calibration curves, cost-benefit curves, and clinical impact curves, the radio-clinical model consistently proved to be the most precise, reliable, and cost-effective model.

Conclusions

We are the first to develop a nomogram based on breast cancer US radiomics features and clinicopathological parameters, that can predict 70-GS risk categorization. It provides an objective and accurate method for patients who are inaccessible to the 70GS.