Year, Author | Sequence | Training and Validation set | Extracted radiomics features, selection, and statistical learning | Biologic correlation and relevance |
---|---|---|---|---|
2008, Diehn | T1, T1+ T2 | T, 22 pts V, 110 pts | -10 binary imaging traits (enhancement, necrosis, mass effect, T2 edema, cortical involvement, SVZ involvement, C:N ratio, contrast/T2 ratio, T2 edema, T2 heterogeneity). - Unsupervised hierarchical clustering, Spearman rank-correlation coefficient. | - Associations between angiogenesis, tumor hypoxia, and the contrast enhancement imaging phenotype; proliferation gene-expression signature and mass effect phenotype; EGFR protein overexpression and contrast:necrosis imaging trait. |
2011, Zinn | FLAIR | T, 26 pts V, 26 pts | - Quantitative models of edema/invasion, enhancing tumor, necrosis. - Comparative marker selection, ingenuity pathway analysis. | - Imaging traits associated with upregulation of mRNA involved in cellular migration/invasion (PERIOSTIN),which was seen to correlate with decreased survival. |
2014, Rahman | ADC-/+ T2/FLAIR | T, 91 pts | - 6 variables extracted from histograms of apparent diffusion coefficient were measured at three times (baseline, post-treatment and change). - Cox proportional hazards model adjusted for clinical variables. | - ADC histogram analysis within both enhancing and nonenhancing components of tumor can be used to stratify for PFS and OS in patients with recurrent glioblastoma treated with Bevacizumab. |
2014, Jamshidi | T1, T1+ T2 Flas | T, 23 pts | - (1) infiltrative versus edematous T2 abnormality, (2) degree of contrast enhancement, (3) necrosis, (4) supraventricular zone (SVZ) involvement, (5) mass effect, and (6) contrast-to-necrosis ratio. - Resampling statistics, analysis of variance, Pearson correlation coefficient. | - Gene-to-trait associations were found such as contrast-to-necrosis ratio with KLK3 and RUNX3, SVZ involvement with the Ras oncogene family and the metabolic enzyme TYMS, and vasogenic edema with the oncogene FOXP1 and PIK3IP1. |
2015, Lee | T1+Flair | T, 65 pts | - 36 spatial habitat diversity (regions with distinctly different intensity characteristics) features based on pixel abundances w/in ROIs. - Overall coefficient of variation, symbolic regression method. | - Association with OS and EGFR+ status - Could be a useful prognostic tool for MRIs of patients with glioblastomas. |
2016, Kickingereder | T1, T1+ Flair | T, 112 pts V, 60 pts | - 4842 total - 17 first-order features, 9 volume and shape features, 162 texture features. - Supervised principal component analysis, Cox proportional hazard models, integrated Brier scores. | - An 72-feature radiomics-based classification of recurrent glioblastoma permits the prediction of treatment outcome to antiangiogenic therapy through PFS and OS. |
2016, Kickingereder | T1+ Flair | T, 79 pts V, 40 pts | - 12,190 indexes - Supervised principal component analysis. | - An 11-feature radiomic signature allowed prediction of PFS and OS, stratification of patients with newly diagnosed glioblastoma, and improved performance compared with that of established clinical and radiologic risk models. |
2016, Grossmann | T1+ FLAIR | T, 144 pts (gene, 91 pts) | - Volumetric features such as the necrotic core, contrast enhancement, abnormal tumor volume, tumor-associated edema, and total tumor volume (TV), as well as ratios of these tumor components. - Spearman rho, C-index, Noether test. | - Association of imaging features with immune response pathways and apoptosis, signal transduction and protein folding processes, homeostasis and cell cycling pathways, as well as OS. |
2016, McGarry | T1, T1+ ADC FLAIR | T, 81 pts | - Map containing 81 (34) potential voxel-wise codes. A 4-digit code was assigned to each voxel. The digit order chosen was T1, ADC, T1+, and FLAIR. Codes ranged from 1111 (dark voxels on all images) to 3333. - Log-rank Kaplan-Meier survival analysis, Cox proportional hazards model, combined classifier. | - Radiomic signature predicted poorer prognosis at tumor diagnosis in newly diagnosed glioblastoma |
2017, Prasanna | T1 FLAIR T2 | T, 65 pts | - 402 radiomic features were obtained for each region: enhancing lesion, peritumoral brain zone and tumor necrosis. - Redundancy maximum relevance feature selection , random forest (RF) classifier, threefold cross-validation. | - Ten radiomic “peritumoral” MRI features, suggestive of intensity heterogeneity and textural patterns, were predictive of survival on treatment-naïve pre-operative glioblastoma. |
2017, Yu | FLAIR | T, 110 pts V, 30 pts | - 671 high-throughput features were extracted from grade II glioma. - Classification by support vector machine and AdaBoost, leave-one-out cross-validation. | - 110 features were selected for the noninvasive IDH1 status estimation of grade II glioma. |
2017, Xi | T1, T1+ T2 | T, 98 pts V, 20 pts | - 1665 imaging features - Reduced using LASSO regularization, classification by support vector machine. | - The best classification system for predicting MGMT promoter methylation status in preoperative gliobastoma originated from the combination of 36 T1, T2, and enhanced T1 images features. |
2017, Kickingereder | T1, T1+ FLAIR T2 | T, 120 pts V, 60 pts | - 1043 imaging features - Penalized Cox model with 10-fold cross-validation. | - The 8-feature radiomic signature increased the prediction accuracy for PFS and OS beyond the assessed molecular, clinical, and standard imaging parameters in newly diagnosed glioblastoma prior to standard-of-care treatment. |
2017, Li | T1+ | T, 96 pts | - 555 imaging features - Student’s tests (t test) | - Glioblastoma in different age groups (< 45 and ≥ 45 years old) present different radiomics-feature patterns, suggesting different pathologic, protein, or genic origins. - 101 features showing the consistency with the age groups, and unsupervised clustering results of those features also show coherence with the age difference. |
2017, Grossmann | T1+ FLAIR | T, 126 pts V, 165 pts | - 65 imaging features from T1 and FLAIR scans at baseline (pretreatment) and follow-up after 6 weeks (post treatment initiation) - Unbiased unsupervised feature selection (PCA), selection of variant features (coefficient of variation). - Minimal redundancy maximal relevance algorithm, Cox proportional hazards model for PFS or OS. | - Multivariable analysis of features derived at baseline imaging resulted in significant stratification of OS and PFS. - These stratifications were stronger compared with clinical or volumetric covariates prognostic value for survival and progression in patients with recurrent glioblastoma receiving bevacizumab treatment. |
2017, Kanas | T1+ FLAIR | T, 86 pts | - 10 quantitative variables and 24 qualitative variables were calculated from the volumes of three distinct regions: edema/invasion, tumor enhancement (tumor), and necrosis. - Isometric feature mapping, locally linear embedding, Laplacian eigenmaps, linear discriminant analysis, factor analysis, principal components analysis, stochastic proximity embedding, random forest, k-nearest neighbors, Gaussian naive Bayes, and the J48 tree. | - The status of MGMT promoter methylation was predicted with an accuracy of up to 73.6%. - Experimental analysis showed that the edema/necrosis volume ratio, tumor/necrosis volume ratio, edema volume, and tumor location and enhancement characteristics were the most significant variables in respect to the status of MGMT promoter methylation in glioblastoma. |
2010, Drabycz | T1+ T2 FLAIR | T, 59 pts | - 4 visual qualitative texture features (cysts, ring/nodular enhancement, margins, homogeneity), volume, 11 regions/sectors features and space–frequency texture analysis based on the S-transform. - Two-way repeated-measures analysis of variance (ANOVA) tests. | - Ring enhancement assessed visually is significantly associated with unmethylated MGMT promoter status. -Texture features on T2 images assessed by the space–frequency analysis were significantly different between methylated and unmethylated cases. |