基于多模态影像和临床特征融合的乳腺癌预后预测模型建立与验证

Establishment and Validation of Breast Cancer Prognosis Prediction Model Based on Multimodal Imaging and Clinical Feature Fusion

目的构建并验证基于多模态影像和临床特征融合的乳腺癌预后预测模型。方法前瞻性选取2020年1月至2023年3月医院收治的156例乳腺癌患者,按照8∶2随机分为训练集(125例)和验证集(31例)。患者新辅助化疗后行手术切除并统计病理完全缓解(pCR)情况。新辅助化疗前后行高分辨磁共振、乳腺超声、钼靶检查,分析影响乳腺癌患者预后(新辅助化疗后pCR)的因素,构建基于多模态影像和临床特征融合的乳腺癌预后预测模型并进行模型的验证及效能评估。结果训练集有31例达到pCR,验证集有7例达到pCR。Logistic回归分析显示肿瘤分期(OR=5.254,95%CI:2.161~12.769)、多普勒超声后方回声(OR=4.909,95%CI:2.020~11.930)、△表观扩散系数(ADC)(OR=4.419,95%CI:1.818~10.741)、钼靶钙化状态(OR=4.358,95%CI:1.793~10.591)是影响乳腺癌患者预后的因素(P<0.05)。以上述影响因素作为预测变量,建立列线图预测模型,各因素总分范围89~374分,对应风险率范围0.07~0.89。列线图模型验证结果显示C-index指数为0.804(95%CI:0.768~0.841),预测乳腺癌预后的校正曲线趋近于理想曲线(P>0.05)。训练集受试者工作特征(ROC)曲线结果显示:列线图模型预测乳腺癌患者预后的灵敏度77.42%(95%CI:58.46%~89.72%),特异度为86.17%(95%CI:77.153%~92.14%),受试者工作特征曲线曲线下面积(AUC)为0.856(95%CI:0.778~0.939)。验证集ROC曲线结果显示:列线图模型预测乳腺癌患者预后的灵敏度71.43%(95%CI:30.26%~94.89%),特异度为91.67%(95%CI:71.53%~98.54%),AUC为0.872(95%CI:0.795~0.949)。结论基于多模态影像和临床特征融合的乳腺癌预后预测模型可有效预测患者预后,列线图模型预测乳腺癌患者预后效能良好。

Objective To construct and validate a breast cancer prognosis prediction model based on multimodal imaging and clinical feature fusion.Methods 156 breast cancer patients admitted to the hospital from January 2020 to March 2023 were prospectively selected and randomly divided into a training set(125 cases)and a validation set(31 cases)according to 8:2.Patients underwent surgical resection after neoadjuvant chemotherapy and pathological complete response(pCR)was analyzed.Before and after neoadjuvant chemotherapy,high-resolution MRI,breast ultrasound and mammography were performed.The factors affecting the prognosis of breast cancer patients(pCR after neoadjuvant chemotherapy)were analyzed,and the prognosis prediction model of breast cancer based on the fusion of multimodal images and clinical features was constructed,and the model was validated and evaluated.Results In the training set,31 cases achieved pCR,and in the verification set,7 cases achieved pCR.Logistic regression analysis showed tumor stage(OR=5.254,95%CI:2.161-12.769),Doppler echo(OR=4.909,95%CI:2.020-11.930),△apparent diffusion coefficient(ADC)(OR=4.419,95%CI:1.818~10.741),molybdenum calcification status(OR=4.358,95%CI:1.793-10.591)were the prognostic factors of breast cancer patients(P<0.05).Taking the above influencing factors as predictive variables,a diagnostic model was established with a nomogram.The total score of each factor ranged from 89 to 374,and the corresponding risk rate ranged from 0.07 to 0.89.The verification results of the nomogram model showed that the C-index was 0.804(95%CI:0.768-0.841),and the correction curve for predicting breast cancer prognosis was close to the ideal curve(P>0.05).ROC curve results of the training set showed that the sensitivity of the nomogram model to predict the prognosis of breast cancer patients was 77.42%(95%CI:58.46%-89.72%),the specificity was 86.17%(95%CI:77.153%-92.14%),and the AUC was 0.856(95%CI:0.778-0.939).The ROC curve results of the validation set showed that the sensitivity of the nomogram model to predict the prognosis of breast cancer patients was 71.43%(95%CI:30.26%-94.89%),the specificity was 91.67%(95%CI:71.53%-98.54%),and the AUC was 0.872(95%CI:0.7955-0.949).Conclusion The prognosis prediction model of breast cancer based on the fusion of multimodal images and clinical features can effectively predict the prognosis of patients with breast cancer.