基于超声的深度学习影像组学模型早期预测乳腺癌新辅助化疗效果的研究

A deep learning prediction model for early evaluation of treatment response to neoadjuvant chemotherapy based on ultrasound images of breast cancer patients

目的探讨深度学习影像组学模型在乳腺癌患者治疗早期预测新辅助化疗(NAC)效果的可行性。方法回顾性选取2018年1月至2021年6月218例接受NAC治疗的乳腺癌患者,患者均完成NAC治疗并在NAC前和NAC第二疗程后进行超声检查。其中166例患者来自机构1(南京医科大学第一附属医院)构成训练集,在Resnet 50卷积神经网络的基础上构建深度学习预测模型;另52例来自机构2(解放军东部战区总医院)的乳腺癌患者构成独立的外部测试集以验证该模型。建立包含关键临床因素的临床模型,通过ROC曲线下面积(AUC)评估各个模型的分辨度。采用DeLong检验对两个模型预测效能和两名超声医师的主观评价进行比较。结果Resnet 50深度学习模型能够准确预测乳腺癌患者NAC的反应。在训练集和外部测试集中AUC分别为0.923(95%CI=0.884~0.962)和0.896(95%CI=0.807~0.980)。并且在两个数据集中深度学习模型也优于临床模型和两名超声医师对NAC效果的预测(均P<0.05)。此外,在Resnet 50深度模型的辅助下,两位超声医师的预测效能显著提高(均P<0.01)。结论基于NAC前和第二疗程后超声图像构建的深度学习影像组学模型,能够在治疗早期对乳腺癌患者NAC的病理反应进行个体化预测。

Objective To investigate the feasibility of deep learning radiomics model in the prediction of neoadjuvant chemotherapy(NAC)response in breast cancer based on ultrasound images at an early stage.Methods Between January 2018 and June 2021,218 patients with breast cancer who underwent NAC were enrolled in the retrospective study.All patients received a full cycle of NAC before surgery and underwent standard ultrasound examination before NAC and after the second cycles of NAC.Of all the patients,166 patients came from institution 1(the First Affiliated Hospital of Nanjing Medical University)were allocated into a primary cohort.Based on the architecture of Resnet 50 convolutional neural,a deep learning prediction model was built.Further validation was performed in an external testing cohort(n=52)from institution 2(General Hospital of Eastern Theater Command,PLA).The clinical model was constructed using independent clinical variables.To evaluate the predictive performance,areas under the curve(AUCs)of these models and two radiologists were compared by using the DeLong method.Results The Resnet 50 model predicted the response of NAC with accuracy.The deep learning model,achieving an AUC of 0.923(95%CI=0.884-0.962)in the primary cohort and an AUC of 0.896(95%CI=0.807-0.980)in the test cohort,outperformed the clinical model and also performed better than two radiologists′prediction(all P<0.05).Furthermore,the two radiologists achieved a better predictive efficacy(AUC 0.832 and 0.808 for radiologists 1 and 2,respectively)when assisted by the DL model(all P<0.01).Conclusions The deep learning radiomics model is able to predict therapy response in the early-stage of NAC for breast cancer patients,which could guide clinicians and provide benefit for timely treatment strategy adjustment.