Keywords:
Computer applications, Liver, CT-Quantitative, CT, Computer Applications-Detection, diagnosis, Segmentation, Computer Applications-General, Tissue characterisation, Cancer, Cirrhosis
Authors:
L. J. Pavan, R. Faletti, A. Di Chio, M. Gatti, A. Depaoli, F. Gentile, F. Guarasci, S. Fiore, P. Fonio; Turin/IT
DOI:
10.26044/ecr2019/C-0289
Results
Three different radiomic signatures were developed.
The first signature (λ=λmin) used the minimum λ value and was based on 15 features (Table 3).
The second signature (λ=1 SE) was calculated with a λ value at one standard error from λmin and was based on 9 features (Table 4).
The third signature (λ=1 SE mod) was derived from the second signature by removing shape features and features from portal venous phase,
therefore considering only 6 features from delayed phase (Table 5).
The signatures showed good discrimination between case and control patients with Area Under Curve (AUC)=0.96 (0.85-1.00),
p<0.001,
SE=85%,
SP=100% for the first signature; AUC=0.94 (0.80-1.00),
p<0.001,
SE=80%,
SP=100% for the second signature and AUC=0.89 (0.73-0.95),
p<0.001,
SE=80%,
SP=86% for the third signature.
This last signature has slightly lower diagnostic performances,
but it’s more easily reproducible and comparable since it’s only based on the delayed venous phase with no shape features.
This delayed phase was the most selected by LASSO regression and it's usually helpful in HCC diagnosis[4].