Radiomics Features of 18F-Fluorodeoxyglucose Positron-Emission Tomography as a Novel Prognostic Signature in Colorectal Cancer

Simple Summary Currently, the optimal treatment for colorectal cancer (CRC) is planned on the basis of the results of preoperative imaging studies. Previous studies investigating the impact of radiomics signatures derived from positron-emission tomography (PET) images mainly focused on patients with rectal cancer, who underwent preoperative chemoradiotherapy, and included a relatively small number of patients, without a validation set. The impact of PET-based radiomics signature analysis in patients undergoing curative-intent radical surgery, with or without chemotherapy, has not been extensively investigated. Thus, we aimed to identify the prognostic value of radiomics signature from18F-fluorodeoxyglucose (18F-FDG) PET images by assessing the imaging features to predict the progression-free survival in patients with CRC. This study demonstrated that radiomics features derived from PET-CT images can help stratify patient prognosis and additionally increase diagnostic accuracy with respect to conventional clinicopathological data-driven prediction model in patients with CRC. Abstract The aim of this study was to investigate the prognostic value of radiomics signatures derived from 18F-fluorodeoxyglucose (18F-FDG) positron-emission tomography (PET) in patients with colorectal cancer (CRC). From April 2008 to Jan 2014, we identified CRC patients who underwent 18F-FDG-PET before starting any neoadjuvant treatments and surgery. Radiomics features were extracted from the primary lesions identified on 18F-FDG-PET. Patients were divided into a training and validation set by random sampling. A least absolute shrinkage and selection operator Cox regression model was applied for prognostic signature building with progression-free survival (PFS) using the training set. Using the calculated radiomics score, a nomogram was developed, and its clinical utility was assessed in the validation set. A total of 381 patients with surgically resected CRC patients (training set: 228 vs. validation set: 153) were included. In the training set, a radiomics signature labeled as a rad_score was generated using two PET-derived features, such as gray-level run length matrix long-run emphasis (GLRLM_LRE) and gray-level zone length matrix short-zone low-gray-level emphasis (GLZLM_SZLGE). Patients with a high rad_score in the training and validation set had a shorter PFS. Multivariable analysis revealed that the rad_score was an independent prognostic factor in both training and validation sets. A radiomics nomogram, developed using rad_score, nodal stage, and lymphovascular invasion, showed good performance in the calibration curve and comparable predictive power with the staging system in the validation set. Textural features derived from 18F-FDG-PET images may enable detailed stratification of prognosis in patients with CRC.

. Patient inclusion of this study. Figure S2. Comparison of heat map of 47 features from all patients before and after standardization. Graphical representation of Heat map of 47 parameters before standardization (A) and after standardization (B). Standardization is the process of putting different variables on the same scale. To standardize variables, the mean and standard deviation for a variable should be calculated. Standardized value could be defined by subtracting the mean and dividing by the standard deviation for each observed value of the variable. This process produces standard scores that represent the number of standard deviations above or below the mean that a specific observation falls.
(A) Ten time cross validation for tuning parameter selection in the LASSO Cox regression model (B) LASSO coefficient profiles of the 47 PET derived features Figure S3. Selection of radiomics signature in PET using LASSO Cox regression model in the training set and definition of rad_score. The least absolute shrinkage and selection operator method (LASSO) was used for regression of high dimensional predictors. The method uses an L1 penalty to shrink some regression coefficients to exactly zero. (A) The partial likelihood deviance (PLD) curve was plotted versus log (λ), where λ is the tuning parameter. Solid vertical lines represent PLD ± standard error (SE). The dotted vertical lines are drawn at the optimal values by using the minimum criteria and 1-SE criteria. Tuning parameter (λ) selection in the LASSO model used 10-fold cross-validation via minimum criteria. A value λ = 0.03577575 with log (λ) = -3.330485 was chosen. (B) LASSO coefficient profiles of the 47 PET derived features. A coefficient profile plot was produced against the log (λ) sequence. The optimal tuning parameter resulted in two non-zero coefficients. Two features, Gray Level Run Length Matrix_Long-Run Emphasis (GLRLM_LRE) and Grey-Level Zone Length Matrix_Short-Zone Low Gray-level Emphasis (GLZLM_SZLGE), with coefficients 0.07079258, 0.11149516 respectively, were selected in the LASSO Cox regression model. The rad_score was defined as 0.07079258 × GLRLM_LRE + 0.11149516 x GLZLM_SZLGE.
A B Figure S4. Cut-off value selection using X-tile plots of the rad_score. (A) X-tile plots of the rad_score and the points of the rad_score coloration of the plot represents the strength of the association at each division ranging from low (dark, black) to high (bright, red or green). Red represents an inverse association between the expression levels and survival of the feature, whereas green represents a direct association. (B) The optimal cut-off value was defined as the value that produced the largest χ2 in the Mantel-Cox test and this point was set as 0.07. Patients were divided into the high-and low-risk subgroups based on this value.
(C) Metabolic tumor volume (MTV) Figure S5. Correlation between rad_score and PET derived conventional parameters such as SUVmax, TLG and MTV. (A) Correlation between SUVmax and rad_score. (B) Correlation between TLG and rad_score. (C) Correlation between MTV and rad_score. The relationship between variables was evaluated using the Spearman rank correlation test.
Note: These mathematic formulas used in this section are mainly derived from the website http://www.lifexsoft.org (accessed on 1 July 2019).
CONVENTIONAL_SUVmean: average SUV in the volume of interest.
CONVENTIONAL_SUVstd: standard deviation SUV in the volume of interest.
CONVENTIONAL_SUVmax: maximum SUV in the volume of interest.

CONVENTIONAL_SUVmax = (4)
CONVENTIONAL_SUVpeak: mean of SUV in a sphere with a volume of ~1 mL and located so that the average value in the VOI is maximum.
CONVENTIONAL_TLG: the product of SUVmean by Volume in mL. HISTO_Energy SHAPE_Volume_mL SHAPE_Volume_vx SHAPE_Sphericity SHAPE_Compacity HISTO_Skewness: asymmetry of the grey-level distribution in the histogram.
HISTO_Skewness : where HISTO (i) corresponds to the number of voxels with intensity i，E is the total number of voxels in the VOI and ̅̅̅̅̅̅̅̅̅ is the average of grey-levels in the histogram.
HISTO_Kurtosis: shape of the grey-level distribution (peaked or flat) relative to a normal distribution.
where HISTO(i) corresponds to the number of voxels with intensity i, E the total number of voxels in the VOI and ̅̅̅̅̅̅̅̅̅ the average of grey-levels in the histogram. HISTO_Entropy_log10: the randomness of the distribution.
where p(i) is the probability of occurrence of voxels with intensity i and =2e-16 HISTO_Entropy_log2: the randomness of the distribution.
where p(i) is the probability of occurrence of voxels with intensity i and =2e-16 HISTO_Energy: the uniformity of the distribution.

HISTO_Energy = ∑ ( )
SHAPE_Volume (mL and voxels): the volume of interest in mL and in voxels.
where Vi correspond to the volume of voxel i of the VOI. SHAPE_Sphericity: how spherical a volume of interest is. Sphericity is equal to 1 for a perfect sphere.

SHAPE_Sphericity
where V and A correspond to the volume and the surface of VOI based on the Delaunay triangulation.
SHAPE_Compacity: how compact the volume of interest is.

SHAPE_Compacity =
where V and A correspond to the volume and the surface of the VOI based on the Delaunay triangulation. Table S3. Definition of radiomics features for Grey level co-occurrence matrix (GLCM).
GLCM_Homogeneity. GLCM_Energy GLCM_Contrast GLCM_Correlation GLCM_Entropy_log10 GLCM_Entropy_log2 GLCM_Dissimilarity The GLCM takes into account the arrangements of pairs of voxels to calculate textural indices. The GLCM is calculated from 13 different directions in 3D with aδ-voxel distance (‖ ⃗ ‖ relationship between neighboured voxels. The index value is the average of the index over the 13 directions in space (X, Y, Z). Seven textural indices can be computed from this matrix. An entry (i,j) of GLCM for one direction is equal to: GLCM_Energy: also called uniformity or second angular moment, the uniformity of grey-level voxel pairs.

GLCM_Correlation = Average over 13 (or 4) directions (∑ ∑
where or corresponds to the average on row i or column j and ℴ and ℴ correspond to the variance on row i or column j.
GLCM_Entropy_log10: the randomness of grey-level voxel pairs.
GLRLM_SRE : Short-Run Emphasis GLRLM_LRE : Long-Run Emphasis GLRLM_LGRE : Low Gray-level Run Emphasis GLRLM_HGRE : High Gray-level Run Emphasis GLRLM_SRLGE : Short-Run Low Gray-level Emphasis GLRLM_SRHGE : Short-Run High Gray-level Emphasis GLRLM_LRLGE : Long-Run Low Gray-level Emphasis GLRLM_LRHGE : Long-Run High Gray-level Emphasis GLRLM_GLNUr : Gray-Level Non-Uniformity for run GLRLM_RLNU : Run Length Non-Uniformity GLRLM_RP : Run Percentage The GLRLM gives the size of homogeneous runs for each grey level. This matrix is computed for the 13 different directions in 3D (4 in 2D) and for each of the 11 texture indices derived from this matrix, the 3D value is the average over the 13 directions in 3D (4 in 2D). The element (i,j) of GLRLM corresponds to the number of homogeneous runs of j voxels with intensity i in an image and is called GLRLM(i,j) thereafter. GLRLM_SRE, GLRLM_LRE: the distribution of the short or the long homogeneous runs in an image. ) (32 )   Table S5. Definition of radiomics features for Neighborhood Grey-Level Different Matrix (NGLDM).

NGLDM_Coarseness
NGLDM_Contrast NGLDM_Busyness The NGLDM corresponds to the difference of grey-level between one voxel and its 26 neighbours in 3 dimensions (8 in 2D). An element (i,1) of NGLDM corresponds to the probability of occurrence of level i and an element (i,2) is equal to: where ̅ (p,q) is the average of intensities over the 26 neighbor voxels of voxel(p,q). NGLDM_Coarseness: the level of spatial rate of change in intensity.

NGLDM_Coarseness =
where E corresponds to the number of voxels in the VOI and G the number of grey-levels. NGLDM_Busyness: the spatial frequency of changes in intensity. GLZLM_SZE : Short-Zone Emphasis GLZLM_LZE : Long-Zone Emphasis GLZLM_LGZE : Low Gray-level Zone Emphasis GLZLM_HGZE : High Gray-level Zone Emphasis GLZLM_SZLGE : Short-Zone Low Gray-level Emphasis GLZLM_SZHGE : Short-Zone High Gray-level Emphasis GLZLM_LZLGE : Long-Zone Low Gray-level Emphasis GLZLM_LZHGE : Long-Zone High Gray-level Emphasis GLZLM_GLNUz : Gray-Level Non-Uniformity for zone GLZLM_ZLNU : Zone Length Non-Uniformity is GLZLM_ZP : Zone Percentage The GLZLM provides information on the size of homogeneous zones for each greylevel in 3 dimensions. Element (i,j) of GLZLM corresponds to the number of homogeneous zones of j voxels with the intensity i in an image and is called GLZLM(i,j) thereafter.