﻿ 基于双参数磁共振成像影像组学机器学习的前列腺癌风险分层
 第二军医大学学报  2021, Vol. 42 Issue (3): 233-242 PDF

Risk stratification of prostate cancer based on biparametric magnetic resonance imaging radiomics machine learning
XING Peng-yi, SONG Tao, WANG Tie-gong, CHEN Lu-guang, MA Chao, YANG Qing-song, LU Jian-ping
Department of Radiology, Changhai Hospital, Naval Medical University (Second Military Medical University), Shanghai 200433, China
Key words: prostatic neoplasms    magnetic resonance imaging    radiomics    machine learning    risk assessment

Gleason分级是前列腺癌进展和生存的最有力预测指标之一，是判断前列腺癌治疗方式的决定因素[7]。Gleason评分系统通过分析腺体分化的程度对前列腺进行组织学识别，然后依据主要肿瘤组织评分及次要肿瘤组织评分判断肿瘤的异质性。Gleason评分3+4分和4+3分的患者有不同的生物学特性，Gleason评分4+3分的前列腺癌患者病死率是Gleason评分3+4分患者的3倍[8]

1 资料和方法 1.1 临床资料

1.2 检查设备与扫描参数

1.3 前列腺病理结果和风险分层

1.4 前列腺癌影像组学分析 1.4.1 MRI图像预处理

1.4.2 MRI病灶分割

1.4.3 影像组学特征提取与选择

1.4.4 模型构建

1.5 统计学处理

2 结果 2.1 临床模型

2.2 影像组学模型

 图 1 前列腺癌MRI影像组学特征提取 Fig 1 MRI radiomics feature extraction of prostate cancer A: Curve of binomial deviation of MRI radiomics model varying with parameter λ; B: Curve of radiomics characteristic coefficient of MRI radiomics model varying with parameter λ; C: Radiomics features screened by MRI radiomics model; D: Comparison of radiomics score (Radscore) between low-risk (Gleason score≤3+4) and high-risk (Gleason score≥4+3) prostate cancer patients of MRI radiomics model training set (left) and validation set (right). The vertical dotted lines in Fig 1A and 1B represent the optimal value of the super parameter λ. In this case, the model and data show the best fit. The numbers above Fig 1A and 1B represent the number of features. As shown in the figure, when the best λ value is obtained, the nearest feature number is 10. The logarithmic function in Fig 1A, 1B and 1C is natural logarithm. The horizontal lines in Fig 1D represent the cut-off value of Radscore (0.43). MRI: Magnetic resonance imaging.

2.3 临床-影像组学联合模型

 图 2 临床-影像组学联合模型对前列腺癌风险分层的诺模图 Fig 2 Nomograph of risk stratification of prostate cancer by clinical-radiomics combined model PI-RADS: Prostate imaging reporting and data system; PSA: Prostate-specific antigen; Radscore: Radiomics score.

 图 3 临床-影像组学联合模型和临床模型ROC曲线对比 Fig 3 Comparison of ROC curves between clinical-radiomics combined model and clinical model A: Training set; B: Validation set. ROC: Receiver operating characteristic; AUC: Area under curve; CI: Confidence interval.

 图 4 临床-影像组学联合模型预测前列腺癌风险分层与病理结果一致性的校准曲线 Fig 4 Calibration curves of clinical-radiomics combined model for predicting the consistency of risk stratification and pathological results of prostate cancer A: Training set; B: Validation set.

 图 5 临床-影像组学联合模型和临床模型决策曲线 Fig 5 Clinical-radiomics combined model and clinical model decision curves

3 讨论