Objective:

This study aimed to preliminarily explore the feasibility of using mag

netic resonance imaging (MRI) three-dimensional texture (3D-texture) features combined with clinical parameters to build predictive models to assess early recurrence of hepatocellular carcinoma (HCC) patients before surgery.

Methods:

A retrospective study of 98 patients with early recurrence of HCC was performed

and laboratory examinations

MRI images

and 3D-texture were recorded. Data redundancy and Lasso regression were used to extract the main features

using a supervised learning algorithm. Modeling was performed and the model was further used to predict early postoperative recurrence in 83 prospective patients.

Results:

The arterial phase (AP) and the portal venous phase (PVP) extracted 6 and 2 texture features respectively for modeling. The prediction performance of the AP-3D-texture model in the training set

validation set and test set was as follows: The accuracy (ACC) of the prediction was 0.735

0.735 and 0.651; Area under the receiver operating characteristic (ROC) curve (AUC) was 0.759

0.769 and 0.669. The PVP-3Dtexture model prediction performance in three data sets was as follows: Training set ACC was 0.721

AUC was 0.591; Validation set ACC was 0.367

AUC was 0.498; Test set ACC was 0.402

AUC was 0.560. The predictive efficacy of AP-3D-texture in combination with clinical parameters in the three data sets was as follows: Training set ACC was 0.838

AUC was 0.876; Validation set ACC was 0.833

AUC was 0.864; Test set ACC was 0.663

AUC was 0.656.

Conclusion:

AP-3D-texture can be used as a predictor of early recurrence of HCC. Combined with clinical parameters

predictive power is further improved

but the efficacy in the test set is low

which is related to the small sample size of the training set.