The need for better patient stratification
The imaging pathway at staging
A role for radiomics?
Feature-group | Parameter examples |
---|---|
First-order-histogram statistics | Mean, median, skewness, kurtosis, energy (uniformity), entropy |
Second-order gray-level co-occurrence matrix (GLCM) statistics | Entropy, homogeneity, energy (uniformity), contrast, autocorrelation, cluster shade, cluster prominence, difference entropy, difference variance, dissimilarity, inverse difference moment, maximum probability, sum average, sum entropy, sum variance |
Second-order gray-level difference matrix (GLDM) statistics | Mean, entropy, variance, contrast |
High-order neighborhood gray-tone difference matrix (NGTDM) statistics | Coarseness, contrast, busyness, complexity, texture strength |
High-order gray-level run-length (GLRL or RLM) statistics | Short run emphasis, long run emphasis, gray-level nonuniformity, run-length nonuniformity, intensity variability, run-length variability |
High-order gray-level size zone matrix (GLSZM) statistics | Short-zone emphasis, long-zone emphasis, intensity nonuniformity, zone percentage, intensity variability, size zone variability |
Fractal analysis | Mean fractal dimension, standard deviation, lacunarity, Hurst component |
18F-FDG PET radiomics
Author | PET time point | Therapy | Features assessed | Outcome and methods | Findings |
---|---|---|---|---|---|
Tixier et al. [26] n = 41 SCC:31 AC:10 | Pre | CRT: 60 Gy with cisplatin or carboplatin/fluorouracil | 38 features including: First order statistics GLCM RLM GLSZM NGTDM | Response prediction: AUROC | Tumor GLCM homogeneity, GLCM entropy, RLM intensity variability and GLSZM size zone variability can differentiate non-responders, partial responders, and complete-responders with higher sensitivity (76%–92%) than any SUV measurement |
Beukinga et al. [27] n = 97 AC:88 SCC:9 | Pre | CRT: 41.4 Gy with carboplatin/paclitaxel | 88 features including: First order statistics Geometry GLCM NGTDM | Response prediction: Models constructed with least absolute shrinkage and selection operator regularized logistic regression | 18F-FDG long run low gray level emphasis higher in responders than non-responders Model including histologic type, clinical T stage, 18F-FDG long run low gray level emphasis and non-contrast CT run percentage has higher AUC than SUVmax: 0.74 vs. 0.54 (after internal validation) |
Nakajo et al. [28] n = 52 SCC | Pre | CRT: 41–70 Gy with cisplatin/5-flurouracil | GLCM: Entropy, homogeneity, dissimilarity; GLSZM: Intensity variability, Size-zone variability, zone percentage | Response prediction: AUROC Prognostication: Multivariable cox analysis | GLSZM intensity variability and GLSZM size-zone variability predictive of response No texture parameter independently associated with progression free or overall survival |
Paul et al. [29] n = 65 SCC:57 AC:8 | Pre | CRT: 50 Gy with platinum chemotherapy & 5-flurouracil | 84 features including: First order statistics GLCM GLSZM GLDM | Response prediction Prognostication: Multivariable cox analysis | Best subset of predictive variables: metabolic tumor volume, GLCM homogeneity Best subset of prognostic variables: WHO performance status, nutritional risk index, metabolic tumor volume |
Foley et al. [17] n = 403 AC:237 + 79 SCC:65 + 22 | Pre | NACT, NACRT, dCRT: not specified | First order statistics GLCM: homogeneity, entropy, dissimilarity; coarseness; GLSZM: intensity variability, large area emphasis, zone percentage; | Prognostication: Multivariable cox analysis | TLG, histogram energy and histogram kurtosis are independently associated with overall survival |
Tan et al. [31] n = 20 AC:17 SCC:3 | Pre-post | CRT: 50.4 Gy with cisplatin/fluorouracil | 192 features including: First order statistics GLCM | Response prediction: AUROC | SUVmean decline, SUV skewness, GLCM inertia, GLCM correlation, and GLCM cluster prominence are predictors of complete response with AUC 0.76–0.85 |
Van Rossum et al. [32] n = 217 AC | Pre-post | CRT: 45 or 50.4 Gy with fluoropyrimidine and either a platinum compound or taxane | 86 features including: First order statistics Geometry GLCM NGTDM | Response prediction: Multivariable Cox analysis | Feature selection by univariable logistic regression Model including induction chemotherapy, change in RLM run percentage, change in GLCM entropy, and post –CRT roundness is better than clinical prediction model |
Yip et al. [33] n = 45 AC:44 SCC:1 | Pre-post | CRT: 45–50.4 Gy with cisplatin, 5-flurouracil, irinotecan/paclitaxel or carboplatin/paclitaxel | GLCM: homogeneity, entropy RLM: high gray run emphasis, short-run high gray run emphasis GLSZM: high gray zone emphasis, short-zone high gray emphasis | Response prediction: AUROC | Response prediction: Change in run length and size zone matrix parameters differentiates non-responders from partial/complete responders (AUC: 0.71–0.76) |
Beukinga et al. [33] n = 70 AC:65 SCC:8 | Pre-post | CRT: 41.4 Gy in 23 fractions with carboplatin/paclitaxel | 113 features including: First order statistics Geometry Local intensity GLCM GLSZM NGTDM | Response prediction: Models constructed with least absolute shrinkage and selection operator regularized logistic regression | Prediction model composed of clinical T-stage and post-NCRT joint maximum adds important information to the visual PET/CT evaluation of a pathologic complete response |
CT radiomics
Author | CT time point & type | Therapy | Features assessed | Outcome and methods | Findings |
---|---|---|---|---|---|
Hou et al. [35] n = 49 SCC | Pre Contrast enhanced | CRT: 60 Gy with nedaplatin/docetaxel or nedaplatin/paclitaxel | 214 features including: First order statistics Geometry GLCM RLM GLSZM NGTDM with Gabor transformation or Gaussian filtration | Response prediction Feature selection: AUROC Prediction model: support vector machine and artificial neural network | Features discriminating non-responders from responders: skewness, kurtosis, GLSZM long zone emphasis, Gabor_MSA-54, Gabor2D_MSE-54 |
Ganeshan et al. [36] n = 21 AC:14 SCC:7 | Pre Non-contrast from PET/CT | No information available | First order statistics with Gaussian filtration | Prognostication: Kaplan–Meier analysis | High uniformity is an independent predictor of survival. Lower uniformity is associated with a poorer overall survival |
Yip C. et al. [37] n = 36 AC:9 SCC:26 | Pre-post Contrast enhanced Arterial Portal venous | CRT: 50 Gy with cisplatin/5-flurouracil or single agent platinum/5-flurouracil | First order statistics with Gaussian filtration | Prognostication: Kaplan-Meier analysis Cox analysis | Higher post treatment entropy (medium/coarse) independently associated with poorer overall survival |
Author | CT time point & type | Therapy | Features assessed | Outcome and method | Findings |
---|---|---|---|---|---|
Ba-Ssalamah et al. [38] n = 67 (Art) AC:47 GIST:15 Lymphoma:5 n = 73 (PV) AC:48 GIST:17 Lymphoma:5 | Pre Contrast enhanced Arterial Portal venous | Not applicable | First order statistics RLM GLCM Absolute gradient Autoregressive model, wavelet transformation | Classification: linear discriminant analysis | Classification of lymphoma vs. AC or GIST feasible on arterial phase: AC vs. lymphoma: 3.1% misclassification GIST vs. lymphoma: 0% misclassification on arterial CT Classification of AC vs. GIST feasible on venous phase: 10% misclassification |
Ma et al. [39] n = 70 AC:40 Lymphoma:30 | Pre Contrast enhanced Portal venous | Not applicable | First order statistics Geometry Texture analysis Wavelet transformation | Feature selection: LASSO Classification: AUROC | 183 radiomic signature identified with potential to differentiate adenocarcinoma from lymphoma Model including histogram root mean squared, GLCM sum variance and absence of peristalsis: AUC 0.86; Sensitivity 70%, Specificity 100%, Accuracy 87% |
Liu et al. [40] n = 107 AC:84 Signet Ring:5 Mucinous:3 Mixed:15 | Pre Arterial Portal venous | Surgery | First order statistics | Classification: AUROC | Arterial phase standard deviation and entropy; portal venous phase mean, max, mode, percentiles are predictive of poorer differentiation |
Yoon et al. [41] n = 26 AC:25 Signet Ring:1 | Pre Portal venous | Trastuzumab-based chemotherapy | First order statistics: Kurtosis, Skewness GLCM: Angular second moment, contrast, entropy, variance, correlation | Prognostication: AUROC Kaplan-Meier | Lower contrast, variance and higher correlation are associated with poorer survival with AUC of 0.77, 0.75 and 0.77 respectively |
Giganti et al. [42] n = 56 AC:37 Signet Ring:19 | Pre Arterial | Surgery | 107 features including: First order statistics GLCM RLM Geometry with Gaussian filtration | Prognostication: Kaplan-Meier, Multivariable Cox analysis | Energy, entropy, skewness are associated with poorer prognosis |
Giganti et al. [43] n = 34 AC:25 Signet Ring:6 | Pre Arterial | NACT: cisplatin/epirubicin/adriamycin/fluoruracil or cisplatin/epirubicin/aadriamycin/capecitabine | First order statistics GLCM Geometry | Response prediction: Multivariable logistic model | Entropy and compactness are higher in responders and uniformity is lower in responders |
MRI radiomics
Author | MRI time point & sequence | Therapy | Features assessed | Outcome and methods | Findings |
---|---|---|---|---|---|
Liu et al. [44]a n = 80 AC:57 Signet ring:10 Mucinous:1 Mixed:12 | Pre ADC-map | Surgery | First order statistics | Staging: Prediction of tumor & nodal stage: AUROC analysis | ADC histogram analysis may differentiate node positive from node negative disease e.g., Percentile ADC10 has an AUC of 0.79 and sensitivity, specificity and accuracy of 72%, 81% and 74% respectively No ability to differentiate T stage |
Liu et al. [45]a n = 87 AC:60 Signet ring:11 Mucinous:1 Mixed:15 | Pre ADC-map | Surgery | First order statistics | Staging: Prediction of tumor & nodal stage: AUROC analysis | Skewness yields a sensitivity and specificity of 76% and 81%, and an AUC of 0.80 for differentiating node positive from node negative gastric cancers |
Zhang et al. [47]a n = 78 AC:58 Signet ring:11 Mucinous:1 Mixed:8 | Pre ADC-map | Surgery | First order statistics | Classification: AUROC analysis | ADC histogram parameters differ between histological grades but with an AUROC < 0.70 this may not be useful in clinical practice |
Liu et al. [46]a n = 64 AC:45 Signet ring:8 Mucinous:2 Mixed:9 | Pre ADC-map | Surgery | First-order Entropy GLCM Entropy | Classification (Grade): AUROC analysis | First-order entropy may differentiate between gastric cancers with vascular invasion with a sensitivity, specificity, accuracy of 86%, 75%, 81% and AUC of 0.82 |