A radiomic model for noninvasive prediction of PD-L1 and VETC expression in hepatocellular carcinoma using enhanced...

Hepatocellular carcinoma (HCC) is a prevalent malignant tumor that is associated with significant morbidity and mortality. Programmed cell death 1 ligand 1 (PD-L1) and Vessel Encapsulating Tumor Clusters (VETC) are critical biomarkers influencing immune evasion and metastasis, making them pivotal for guiding treatment decisions. However, it is challenging to obtain pathological samples from some patients because of factors such as advanced tumor stage and poor liver function.

This study aimed to develop an AI-based imaging model to non-invasively predict PD-L1 and VETC expression in HCC patients, addressing the challenge of limited histopathological data.

This retrospective study included 162 HCC patients diagnosed between January 2017 and December 2022. Patients were randomly divided into training and test sets (8:2). Radiomic features were extracted from CT images, and various machine learning algorithms were used to construct predictive models and assess their accuracy in predicting PD-L1 and VETC expression.

A total of 2,286 features were extracted from the enhanced abdominal CT images. Among them, seven features were associated with PD-L1 expression and 10 with VETC expression. The Random Forest (RF) model demonstrated good calibration and fit, emerging as the most effective with an AUC of 0.834 (95% CI: 0.752-0.915) for PD-L1 and 0.883 (95% CI: 0.818-0.949) for VETC in the training set, while achieving AUCs of 0.740 (95% CI: 0.541-0.939) for PD-L1 and 0.705 (95% CI: 0.488-0.922) for VETC in the test set.

The radiomics model derived from enhanced abdominal CT demonstrates its potential as a noninvasive tool for predicting the expression of PD-L1 and VETC in HCC patients.