Skip to main content

Table 5 Reports on outcome prediction of SBRT in lung cancer from analysis of radiomic features

From: Radiomics for prediction of radiation-induced lung injury and oncologic outcome after robotic stereotactic body radiotherapy of lung cancer: results from two independent institutions

Author

N

Modality/features (software applied)

# features selected

Model type

Outcome measures

Validation

Result/comment

Huynh [24]

113

CT:1605 (in-house software)

12 + clinical

Survival analysis, cc-index

Recurrence, Distant mets., OS

Single institution cross validation

Risk for recurrence: no significant features Risk for dist. metastases: 1 sign. Feature OS: 4 significant features, cc = 0.67

Li [28, 29]

92

CT: 219 (Definiens Developer)

8–68 + clinical + semantic

ROC-analysis

Recurrence, RFS, OS

Single institution cross-validation

Risk stratification: AUC = 0.69–0.75

Zhang [58]

112

CT: 30 (ProCanVAS)

dependent on model

8 models: Random forest GLM, SVM etc

Recurrence, Distal failure, OS

Single institution cross validation

Risk stratification: AUC = 0.60–0.77

Yu [34]

442

CT: 12 (IBEX)

2

Random survival forests

Regional recurrence, OS

Single institution test set: 67%

OS risk stratification: p = 0.017 Recurrence risk stratification: p < 0.05 2 sign. features: kurtosis, homogeneity

Li [30]

110

CT + FDG-PET (learned by model)

from model

Kernelled support tensor machine

Distant failure

Single institution test set: 30%

Risk stratification: AUC = 0.80

Oikonomou [31]

150

CT + FDG-PET 2 × 21 (ProCanVAS)

6–8, 4 from PCA

PCA, logistic regression

Local control, Distant control, DSS, OS

Single institution cross validation

Risk stratification: p = 0.004–0.02 features: heterogeneity and morphology

Starkov [32]

116

CT: 2D-textures from solid core and GGO

2–30

Cox regression lasso

PFS, distant failure

Single institution cross validation

Risk stratification: p = 0.03 dependent on wavelet filtering

Lafata [26]

70

CT: 43

2

Logistic regression regularized

Local recurrence

none

Risk stratification: p = 0.048 features: density

Franceschini [23]

102

CT: 41 (LifeX)

4–6

Cox regression elastic net, back selection

Nodal relapse, PFS, DSS

Single institution Test set: 32%

Nodal Relapse: accuracy = 85% PFS: 53 vs.45 months features: heterogeneity

Lou [59]

944*

CT

learned by model

CNN, Multivariate competing risk

Local recurrence

Multi institution test set: 10%

Risk stratification: p < 0.002

Baek [21]

122

CT + FDG-PET

2 × 55,296

Features from k-medoids pool

CNN (U-Net) logistic regression

OS

Independent institution test set: 21%

Risk stratification: AUC = 0.87

  1. ProCanVAS prostate cancer visualization and analysis system, PCA principal component analysis, cc-index: concordance index, RFS Recurrence-free survival, ROC receiver-operator-characteristics; validation by independent test sets shown in bold
  2. *Includes recurrent lung cancers and pulmonary metastases
  3. #Features from PET and CT