基于不同感兴趣区的多参数MRI影像组学在前列腺癌侵袭性评估中的应用研究

Application of Multiparametric MRI Imaging Based on Different Regions of Interest in the Assessment of Prostate Cancer Aggressiveness

目的观察在多参数MRI上基于不同感兴趣区的影像组学对前列腺癌侵袭性的评估价值。方法回顾性搜集来自两个不同中心经病理确诊的前列腺癌患者,中心1纳入123例作为训练队列,53例作为内部验证队列。中心2纳入69例患者作为外部验证队列。在多参数MRI上采用两种不同的分割方式:分割前列腺区域和病灶区域,经过特征提取与降维,得到影像特征集合,利用逻辑回归算法,基于3种组学特征集合建立了3种组学模型:模型1(基于前列腺区域),模型2(基于病灶区域),模型3(基于前列腺+病灶)。分析与前列腺癌侵袭性相关的临床特征,建立影像组学特征+临床特征的联合模型(模型4)。用受试者工作特征曲线分析比较每个模型在评估前列腺癌侵袭性方面的表现。结果基于前列腺区域的感兴趣区保留了57个最佳特征,基于病灶区域的感兴趣区保留了27个最佳特征,基于前列腺区域+病灶区域的感兴趣区域保留了64个最佳特征。与仅分割前列腺区域的模型(AUC 0.89)和仅分割病灶区域的模型(AUC 0.87)相比,同时分割前列腺和病灶的模型(AUC 0.92)在评估前列腺癌侵袭性方面表现更好,并且与结合了组学特征和临床特征的联合模型性能(AUC 0.92)相当。结论结合前列腺和病灶区域的影像组学模型在评估前列腺癌侵袭性方面的性能可能优于仅分割病灶区域或仅分割前列腺区域的影像组学模型。

Objective To observe the value of radiomics based on different regions of interest on multiparametric MRI for the assessment of prostate cancer aggressiveness.Methods In this retrospective study,patients with pathologically confirmed prostate cancer from two different centers were included,with center 1 including 123 cases as a training cohort and 53 cases as an internal validation cohort.Center 2 included 69 patients as an external validation cohort.Two different segmentation methods were used on multiparametric MRI:segmentation of prostate region and lesion,and after feature ex⁃traction and reduction,radiological feature sets were obtained.Using logistic regression algorithms,three radiomic models were established based on three radiomic feature sets:model 1(based on prostate region),model 2(based on lesion re⁃gion),and model 3(based on prostate+lesion).Clinical features associated with prostate cancer aggressiveness were ana⁃lyzed to create an integrated model combining radiomic features with clinical features(model 4).Receiver operating charac⁃teristic(ROC)curve analysis was used to compare the performance of each model in assessing prostate cancer aggressive⁃ness.Results The region of interest based on the prostate region retained the 57 best features,the region of interest based on the lesion region retained the 27 best features and the region of interest based on the prostate region+lesion retained the 64 best features.The model with both prostate and lesion region segmented(AUC 0.92)performed better in assessing prostate cancer aggressiveness compared to the model with only prostate region segmented(AUC 0.89)and the model with only lesion region segmented(AUC 0.87),and was comparable to the performance of the integrated model combining ra⁃diomic and clinical features(AUC 0.92).Conclusion A radiomics model combining the prostate region and the lesion region may perform better in assessing prostate cancer aggressiveness than a radiomics model segmenting only the lesion re⁃gion or segmenting only the prostate region.