Table 3.
Author and year of publication | Patient sample size and disease stage | Time of imaging and study type | Therapy | Number of parameters derived | Method of feature grouping/selectionand number of parametrs derived | Outcome parameter | Result |
---|---|---|---|---|---|---|---|
Bogowicz et al16 May 2017 |
TC = 93 ; VC = 56 Stage III and IV HNSCC(Orp,Hyp,Lar,OC) |
Pretreatment Retrospective for TC Prospective for VC |
Definitive IMRT 70 Gy with cisplatin or cetuximab. | Shape, Intensity, Texture, Wavelet transform 317 features |
Grouping: PCA Selection: UVCRA for prognosis (9) For comparison with clinical and combined radiomics-clinical model: MVCRA(3) Split ROC curves at 18 mths |
Predict LC: using CI Compare radiomics versus clinical model and a combined clinicoradiomic model for LC |
Radiomics signature significantly associated with LC Combined radiomics + clinical model performed better than radiomics model alone in TC, but not VC |
Bogowicz et al28 June 2017 |
N = 172 TC = 121; VC = 51 Stage III and IV HNSCC(Orp,Hyp,Lar,OC) |
Pretreatment Retrospective for TC Prospective for VC |
Definitive IMRT 70 Gy with cisplatin or cetuximab TC VC = TC+_ consolidation cetuximab |
Shape, Intensity, Texture, Wavelet transform 569 |
Combination of feature selection using PCA and classification using Cox regression with backward selection: chosen for least complicated and best discriminatory. Model validation: CI using Wilcoxon and bootstrap |
Compare CT, PET, PETCT radiomics models for prognosis | CT radiomics overestimates probability of tumor control in high risk group. Mostly due to CT artifacts and variable contrast dose. CT (GLSZM,HLH) PET (Spherical disproportion, GLSZM) Combined (CT HLH and PET GLSZM) All showed similar discriminatory CI > 0.7 |
El Naqa et al27 June 2009 |
N = 9 Stage and type: NM |
Pretreatment Retrospective Median F/U period of 30 months. |
Chemoradiotherapy (details not mentioned) | IVH, Shape, Texture, SUV measures 18 |
RS and AUC for association between extracted features and post-radiotherapy outcomes. Two-metric logistic regression model |
Analyzed for endpoint of overall survival rate | Shape-based metrics had the highest categorical prediction power, while commonly used SUV descriptive statistics had the lowest predictive ability |
Cheng et al26 Sept 2013 |
N = 70 T3-4 OPSCC Follow up: 24 mths In-house (Matlab) |
Pretreatment Retrospective |
Completed platinum-based CCRT, cetuximab-based CBRT, or RT alone with curative intent | SUV histogram, TLG, NGLCM, NGTDM | MVCRA to identify the independent predictors of PFS, DSS, and OS RS to evaluate the associations between textural characteristics, SUVmax, MTV, TLG, and the general characteristics of the study participants. |
Can textural features provide any additional prognostic information over TLG and clinical staging | Uniformity extracted from the normalized gray-level cooccurrence matrix found to be an independent prognostic predictor |
Cheng et al25 Oct 2014 |
N = 88 T3 or T4 OPSCC In-house (Matlab) |
Pretreatment Retrospective |
83 patients received CCRT, three received BRT, and the remaining two patients received RT alone with curative intent. | SUV, TLG, GLRLM, GLSZM | UV and MVCRA to identify the independent predictors of PFS and DSS. Kaplan-Meier curves for survival. |
Prognostic impact of regional heterogeneity on Progression-free survival (PFS) and disease-specific survival (DSS) | Zone-size nonuniformity (ZSNU) identified as an independent predictor of PFS and DSS. Model combining total lesion glycolysis, uniformity and ZSNU showed a higher predictive value than each variable alone |