The Value of Machine Learning-based Radiomics Model Characterized by PET Imaging with Ga-FAPI in Assessing Microvascular Invasion of Hepatocellular Carcinoma.

RATIONALE AND OBJECTIVES

This study aimed to develop a radiomics model characterized by Ga-fibroblast activation protein inhibitors (FAPI) positron emission tomography (PET) imaging to predict microvascular invasion (MVI) of hepatocellular carcinoma (HCC). This study also investigated the impact of varying thresholds for maximum standardized uptake value (SUV) in semi-automatic delineation methods on the predictions of the model.

METHODS

This retrospective study included 84 HCC patients who underwent Ga-FAPI PET and their MVI results were confirmed by histopathological examination. Volumes of interest (VOIs) for lesions were semi-automatically delineated with four thresholds of 30%, 40%, 50%, and 60% for SUV. Extracted shape features, first-, second- and higher-order features. Eight PET radiomics models for predicting MVI were constructed and tested.

RESULTS

In the testing set, the logistic regression (LR) model achieved the highest AUC values for three groups of 30%, 50%, and 60%, with values of 0.785, 0.896, and 0.859, respectively, while the random forest (RF) model in 40% group obtained the highest AUC value of 0.815. The LR model in 50% group and the extreme gradient boosting (XGBoost) model in 60% group achieved the highest accuracy, each at 87.5%. The highest sensitivity was observed in the support vector machine (SVM) model in 30% group, at 100%.

CONCLUSION

The Ga-FAPI PET radiomics model has high efficacy in predicting MVI in HCC, which is important for the development of HCC treatment plan and post-treatment evaluation. Different thresholds of SUV in semi-automatic delineation methods exert a degree of influence on performance of the radiomics model.