A Study on the Impact of Image Processing Methods on Predicting the MGMT Gene Status of Glioblastomas through Radiomics Analysis
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
万方数据
目的 基于随机森林分类器找到各常规MRI序列及各序列联合后最适合的图像处理方法,并联立各经过最适合的图像处理方式的序列建立一个术前预测胶质瘤MGMT基因甲基化状态的影像组学模型.方法 回顾性搜集了 104例(河北医科大学第二医院数据75例,TCIA公共数据集29例),经病理证实为MGMT基因甲基化(56例)和非甲基化(48例)的胶质瘤患者的T1 WI、T2 WI、FLAIR、CE-T1 WI图像.然后,由两名医师进行感兴趣区勾画,勾画范围包括瘤周水肿区、坏死区、实质区.之后对各序列的感兴趣区分别进行标准分数(Z-score)、Nyúl图像处理,再各自进行灰度离散化处理,并对各序列处理后的图像使用Pyradiomics包进行特征提取.对各序列进行特征筛选并进行随机森林建模,得到各序列最佳图像处理方式.再联合多序列进行特征筛选并建模,得到多序列最佳图像处理方式.最后将经过最佳处理方式的各序列特征联合进行筛选并通过随机森林建模.结果 在单序列中T1 WI配合Nyúl序列表现最佳,训练集及验证集受试者工作特征曲线曲线下面积(AUC)值为0.98,0.85;在多序列建模中,Nyúl图像归一化方法最佳,训练集与测试集AUC值为0.94,0.89;在所有建立的模型中联合经过各序列最佳图像处理方式处理的各序列建立的模型性能最佳,训练集与验证集AUC值为0.98,0.92.结论 基于联合经过各序列最佳图像处理方式处理的各序列建立的模型性能,在多序列模型中,性能最佳;在单序列中,T1 WI适合Nyúl,T2 WI 适合 Z-score&FBN32,FLAIR 适合 Z-score&FBN128,CE-T1 WI 适合 Nyúl&FBS1/128.
Objective Radiomics analysis is often affected by variations introduced by different imaging devices and sites.This study aims to identify the most suitable image processing methods for individual conventional MRI sequences and their combinations using a random forest classifier.Additionally,the study aims to develop a radiomics model for preopera-tive prediction of MGMT gene methylation status in gliomas by combining sequences processed with their optimal image pro-cessing methods.Methods Firstly,a retrospective collection was conducted,involving 104 cases,with 75 cases from He-bei Medical University Second Hospital and 29 cases from The Cancer Imaging Archive(TCIA)public dataset.These ca-ses were pathologically confirmed to be glioma patients with MGMT gene methylation(56 cases)and non-methylation(48 cases).The MRI images included T1 WI,T2 WI,FLAIR,and CE-T1 WI sequences.Subsequently,two physicians delineated regions of interest,encompassing peritumoral edema,necrotic zone,and solid region.The regions of interest for each se-quence were subjected to Z-score and Nyúl image processing.Subsequently,individual sequences underwent grayscale dis-cretization.Pyradiomics package was employed to extract features from the processed images of each sequence.Feature se-lection was performed for each sequence,followed by building random forest models.The optimal image processing methods for each sequence were determined through these models.Then,a combination of multiple sequences underwent feature se-lection and modeling to identify the optimal image processing methods for multi-sequence scenarios.Finally,the features from each sequence,processed with their optimal methods,were combined and used to build a predictive model through ran-dom forest modeling.Results In the single-sequence analysis,T1 WI combined with Nyúl normalization demonstrated the optimal performance,with AUC values of 0.98 for the training set and 0.85 for the validation set.In the context of multi-se-quence modeling,the Nyúl image normalization method proved to be the best,yielding AUC values of 0.94 for the training set and 0.89 for the testing set.Among all the established models,the combined model,integrating features from each se-quence processed with their respective optimal image processing methods,exhibited the best performance,with AUC values of 0.98 for the training set and 0.92 for the validation set.Conclusion This study explored and obtained the optimal im-age processing methods based on random forest for single and multiple sequences.The performance of models established for each sequence,processed with its optimal image processing method,was evaluated.Among these models,the multi-se-quence model demonstrated the best performance.For single sequences,it was found that T1 WI is suitable for Nyúl,T2 WI is suitable for Z-score & FBN32,FLAIR is suitable for Z-score & FBN128,and CE-T1 WI is suitable for Nyúl & FBS1/128.In addition to establishing a well-performing model,this study provides recommendations for image processing for future re-searchers.
NETL
NSTL
万方数据
