Keywords:
Artificial Intelligence, Liver, Abdomen, CT, MR, Neural networks, Diagnostic procedure, Cancer
Authors:
J. Thüring1, O. Rippel1, C. Haarburger1, D. Merhof1, P. Schad1, P. Bruners1, C. K. Kuhl1, D. Truhn2; 1Aachen/DE, 2Cologne/DE
DOI:
10.26044/ecr2019/C-0355
Results
Epidemiologic,
laboratory,
and clinical characteristics are shown in table 2. In total,
11 significant radiomic imaging features were found that correlate significantly with the Child-Pugh Score in univariate analysis.
Spearman’s correlation coefficient was significant for all machine learning algorithms,
albeit strongest for the CNN.
The radiologists’ rating exhibited the strongest correlation (ρLR= 0.35,
ρRF= 0.32,
ρCNN= 0.51,
ρRP= 0.60; all p<0.001).
The accuracy of the CNN and RP was significant better as compared to the no information rate (ACCLR= 47%,
p= 0.47; ACCRF= 47%,
p= 0.38; ACCCNN= 53%,
p=0.03; ACCRP= 57%; p<0,001; NIR= 46%) (Fig.
2).
If low disease severity (Child-Pugh class= A) versus advanced disease severity (Child-Pugh class≥ B) was classified,
only the CNN revealed significant results against the NIR (LR: accuracy= 71%; sensitivity= 85%,
specificity= 38%; p= 0.48; RF: accuracy= 70%; sensitivity= 81%,
specificity= 43%; p= 0.58; CNN: accuracy= 78%; sensitivity= 81%,
specificity= 70%; p< 0.001; RP: accuracy= 71%; sensitivity= 82%,
specificity= 66%; p= 0.53).
The AUC in the ROC was highest for the CNN (AUCCNN= 0.80),
followed by the radiologists’ predictions (AUCRP= 0.76),
the LDA (AUCLR= 0.71) and the RF classifier (AUCRF= 0.69) (Fig.
3).