Out of 202 patients who underwent OLT for HCC in our Centre between 2011 and 2015,
we retrospectively selected 8 (4%) patients (8 males,
mean age 61 years,
total of 20 lesions,
Group 1 Table 1) who had HCC recurrence within two years after OLT and underwent triple-phase CT within 3 months prior to treatment (Lightspeed Pro16 or Optima 660 MG40,
Ge Healthcare,
Table 2).
Eleven age- and sex-matched patients (9 males,
mean age 63 years,
total of 22 lesions,
Group 2 Table 1) who underwent OLT for HCC between 2014 and 2015,
with at least 2-year disease-free follow-up,
served as control group.
Only lesions with a pathological confirmation were analyzed.
Necrotic lesions due to bridge or downstaging treatments before OLT were excluded,
while relapses or new nodules were included in the analysis.
Using IBEX® (Imaging Biomarker Explorer) software[1],
lesions were segmented in a semi-automatic way,
drawing 5 ROIs on axial images and obtaining the missing ROIs to create a volume of the lesion.
Every lesion was segmented in each CT phase by the same operator and elaborated with two preprocessing filters (Buttersworth smoothing and 8-bit depth resample),
in order to reduce Gaussian noise and matrix bin[2].
Finally,
152 texture parameters were obtained for analysis for each CT phase (total of 456 features).
Extrapolated features included: gray level run length matrix (n=11),
intensity features (n=38),
intensity histogram features (n=15),
neighborhood intensity difference matrix 3D (n=5),
shape features (n=17),
gray level co-occurrence matrix 3D (GLCM3) features (n=66).
Categorical variables were analyzed with Fisher test,
while continuous variables with non-parametric Mann-Whitney test.
In order to create a score with the fewest number of variables,
a LASSO (Least Absolute Shrinkable and Selection Operator) logistic regression model was used,
in the Elastic Net version[3].
This tool allowed to create three radiomic signatures (rad-score) with different values of lambda and different analyzed features.
The predictive performance of each radiomic signature was evaluated via receiver operating curve (ROC).
Statistical significance was p<0.05.