多模态超声特征结合机器学习可预测乳腺浸润导管癌中Ki-67的表达水平

目的  探讨多模态超声特征结合机器学习预测乳腺浸润导管癌中Ki-67高表达的可行性。方法  回顾性分析155例乳腺浸润导管癌患者，155个病灶经病理证实。术前行常规超声和声辐射力脉冲成像，免疫组化染色记录Ki-67的表达，将患者分为Ki-67高表达组（n=105）和低表达组（n=50）。采用Logistic回归分析得出独立危险因素，采用随机森林及Logistic回归模型预测。结果  单因素分析显示Ki-67表达与肿块最大径、边界、腋窝淋巴结状态、阻力指数、声触诊组织成像及声触诊组织定量的差异有统计学意义（P < 0.05）。多因素分析结果显示，最大直径、边界、声触诊组织定量及阻力指数对Ki-67为独立危险因素。随机森林模型结果显示，Ki-67表达影响因素的重要性排序依次是最大直径、声触诊组织定量、阻力指数及边界。随机森林及Logistic回归模型预测乳腺浸润导管癌中Ki-67高表达曲线下面积分别为0.871、0.866，Ki-67值与肿块直径呈正相关关系（r= 0.319，P < 0.001）。结论  多模态超声特征结合机器学习可用于预测乳腺浸润导管癌Ki-67的表达水平。

Value of combining multimodal ultrasound features with machine learning to predict high expression of Ki-67 in breast infiltrating ductal carcinoma

Objective To investigate the value of multi- modal ultrasound features combined with machine learning in predicting high expression of Ki-67 in breast invasive ductal carcinoma. Methods A retrospective analysis was conducted in 155 patients with invasive ductal carcinoma and 155 lesions confirmed by pathology. Preoperative conventional ultrasound and acoustic radiation force impulse were performed, immunohistochemical staining was used to record the expression of Ki-67, and the patients were divided into overexpression groups(n=105)and low expression groups(n=50). Logistic regression analysis was used to analyze the differential indicators to obtain independent risk factors, and random forest and Logistic regression models were used for prediction. Results Univariate analysis showed that there were significant differences in the expression of Ki-67 and the maximum diameter, boundary, axillary lymph node status, resistance index, virtual touch tissue imaging and virtual touch tissue quantification of the lesion (P < 0.05). Multivariate analysis showed that the maximum diameter, boundary, virtual touch tissue quantification and resistance index were independent risk factors for Ki-67 expression. The random forest model showed that the influencing factors for Ki-67 expression were ranked in order of importance as the maximum diameter, virtual touch tissue quantification, resistance index and boundary. The areas under the curve of the random forest and logistic regression models in predicting high expression of Ki-67 in breast invasive ductal carcinoma were 0.871 and 0.866, respectively. There was a positive correlation between the expression level of Ki-67 and the diameter of the lesion (r=0.319, P < 0.001). Conclusion Multi-modal ultrasound features combined with machine learning can be used to predict the level of Ki-67 expression in invasive ductal carcinoma, providing reference value for clinical diagnosis and treatment.