多方位MRI影像组学预测局部进展期直肠癌新辅助放化疗疗效

Multi-directional MRI radiomics to predict the efficacy of neoadjuvant chemoradiotherapy for locally advanced rectal cancer

目的:采用MRI影像组学方法,提取局部进展期直肠癌(LARC)病灶影像组学特征,并联合临床及常规影像特征构建预测模型,探讨模型对新辅助放化疗(nCRT)疗效的预测效能。方法:回顾性分析209例LARC患者nCRT前的临床及影像资料。LARC患者在nCRT后6~8周行全系膜切除术(TME),并评估肿瘤病理退缩分级(TRG)。按疗效分为nCRT反应良好组(TRG0~1级)和反应不良组(TRG2~3级),按照1:1随机分为训练组和验证组对模型进行内部验证。手动勾画T 2WI序列横轴面(TRA)、矢状面(SAG)及冠状面(COR)图像提取影像组学特征,采用LASSO回归筛选特征并构建影像组学标签。通过多因素logistics分析筛选nCRT疗效的独立预测因子并构建联合预测模型,采用ROC曲线及校正曲线对模型进行评估,并使用临床决策曲线评价模型的临床价值。结果:209例患者中nCRT反应良好组61例,反应不良组148例。T 2WI序列横轴面、矢状面、冠状面图像各提取379个影像组学特征,ICC为0.9的特征中TRA 96个,SAG 88个,COR 91个。LASSO回归筛选的特征中TRA 4个,SAG 5个,COR 2个,联合图像(COM)7个。TRA、SAG、COR、COM等四个影像组学标签模型预测nCRT疗效的AUC值分别为0.637、0.682、0.619、0.731。多因素logistics分析结果表明COM影像组学标签、升高的CEA(>3.4 ng/mL)和壁外血管侵犯(EMVI)是nCRT疗效的独立预测因子,联合三者构建的联合预测模型预测nCRT疗效的AUC值为0.811,模型得到了很好的内部验证,校准度较高,临床决策曲线显示了较高的临床应用价值。结论:联合T 2WI序列横轴面、矢状面、冠状面图像构建的多方位影像组学标签可以在治疗前对LARC患者nCRT疗效进行预测,联合升高的CEA(>3.4 ngmL)、EMVI、COM影像组学标签等构建的模型,其预测效能进一步提高,可以辅助制定个体化诊疗方案,实现最大治疗收益。

Objective:The radiomics features of locally advanced rectal cancer(LARC)lesions were extracted using the MRI radiomics method,and a prediction model was constructed by combining clinical and conventional imaging features to explore the predictive efficacy of the model on the efficacy of neoadjuvant chemoradiotherapy(nCRT).Methods:The clinical and imaging data of 209 LARC patients before nCRT were retrospectively analyzed.Total mesorectal resection(TME)was performed on the patients with LARC 6~8 weeks after nCRT,and were assessed for tumor pathological regression grading(TRG).According to the efficacy,the patients were divided into a group of good nCRT response(TRG0~1 grade)and a group of poor response(TRG2~3 grade).It were randomly divided into training and verification group in accordance with 1:1 for internal verification of the model.The radiomics features were extracted through mannually sketching the transverse,sagittal and coronal T 2WI images.The features were screened by LASSO regression,and the radiomics labels were constructed.Then,the independent predictors of the efficacy of nCRT were screened by multifactorial logistic analysis to construct prediction model.The ROC curves and correction curves were used to evaluate the model,and the clinical value of the model was evaluated by clinical decision curves.Results:Among 209 patients,61 had good nCRT response and 148 had poor nCRT response.A total of 379 radiomics features were extracted from each of the transverse,sagittal and coronal T 2WI images.The features with ICC≥0.9:TRA:96,SAG:88,COR:91.LASSO regression screening features:TRA:4,SAG:5,COR:2,COM:7.The AUC values of TRA,SAG,COR and COM radiomics labels models in predicting the efficacy of nCRT were 0.637,0.682,0.619 and 0.731,respectively.The results of multivariate logistic analysis showed that the COM radiomics labelling,elevated CEA level(>3.4ng/mL)and extra-mural vascular invasion(EMVI)were independent predictors of the efficacy of nCRT.The area of AUC cureve of prediction model was 0.811.The model was well validated internally,with high calibration,and the clinical decision curve showed high clinical application value.Conclusions:The multi-directional radiomics labels constructed by combining transverse,sagittal and coronal images form T 2WI sequences can predict the efficacy of nCRT in LARC patients.The predictive efficacy of the model is further improved by combining the elevated CEA(>3.4ng/mL),EMVI and COM radiomics label,which can assist in formulating individualized diagnosis and treatment plans to achieve maximum benefits.