Objective:

To eval

uate the performance of a deep learning model based on the Kupffer phase of perflubutane microspheres for injection (product name Sonazoid) contrast-enhanced ultrasound in predicting microvascular invasion (MVI) of hepatocellular carcinoma (HCC)

comparing it with radiomics model and clinical model.

Methods:

This study retrospective included 146 patients with primary HCC who underwent Sonazoid contrast-enhanced ultrasound examination in The First Affiliated Hospital of Guangxi Medical University from July 2020 to September 2022

randomly divided into a training set of 102 and a validation set of 44 in a 7∶3 ratios. Based on the region of interest in tumors

ResNet101 model was used to extract deep learning features through transfer learning

and PyRadiomics was utilized to extract radiomics features. Mann-Whitney

U

test and least absolute shrinkage and selection operator (LASSO) algorithm were employed to reduce features dimension. LASSO regression was used to construct both the deep learning model and radiomics model

a clinical model was also built based on clinical features. The diagnostic performance of models was evaluated using the area under the receiver operating characteristic curve (AUC)

sensitivity

specificity

and accuracy. DeLong testing algorithm was used to compare the diagnosticperformance between models.

Results:

In the training set

the AUC (95% CI) for the deep learning model

radiomics model

clinical model was 0.931 (0.880-0.981)

0.823 (0.744-0.903) and 0.719 (0.614-0.824)

respectively. In the validation set

the AUC (95% CI) for the deep learning model

radiomics model

clinical model was 0.895 (0.757-1.000)

0.711 (0.514-0.909) and 0.606 (0.390-0.822)

respectively. DeLong testing indicated that in both the training and validation sets

the diagnostic performance of the deep learning model was superior to that of the radiomics model and clinical model (

P

＜0.05). Both univariate and multivariate logistic regression analyses showed that A

FP (

P

＜0.05) and Barcelona Clinic Liver Cancer staging (

P

＜0.001) could be used as independent predictors of MVI in HCC patients.

Conclusion:

The deep learning model based on the Kupffer phase of Sonazoid contrast-enhanced ultrasound demonstrates excellent performance in predicting MVI in HCC patients. It has the potential to become a non-invasive imaging biomarker for predicting MVI.