Characterizing the tumor microenvironment non-invasively after sorafenib and memory-like NK cell immunotherapy in hepatocellular carcinoma

Hepatocellular carcinoma (HCC) is a highly heterogeneous malignancy with limited therapeutic options and is often complicated by treatment resistance. Recent evidence suggests that memory-like natural killer ( p NK) cells may offer a potent immunotherapeutic approach, particularly when combined with sorafenib; however, the optimal delivery method and ability to non-invasively monitor treatment effects remain unclear. This study aimed to evaluate the efficacy of sorafenib and p NK cell combination therapy across different administration routes and to develop machine learning models for non-invasive monitoring of the tumor microenvironment (TME). Two distinct rat HCC models (N1–S1 and McA-RH7777) were used to address the heterogeneity of HCC characteristics. At the beginning of tumor xenograft implantation, the animals were randomly assigned to control, sorafenib, p NK, or combination groups. The N1–S1 model received hepatic arterial delivery of p NK cells, whereas the McA-RH7777 tumor model underwent systemic administration via the tail vein. Tumor growth was monitored weekly using quantitative MRI. Histopathological analysis was performed on H&E, CD56, CD31, and TUNEL-stained tissue slides to evaluate the TME. Quantitative texture features were extracted from multiparametric MRI data and used to train multiple radiofrequency regression models to predict histopathological changes. The N1–S1 tumor model demonstrated superior responsiveness to all therapies compared with the McA-RH7777 model. Histopathological analysis revealed significant baseline differences, with N1–S1 tumors exhibiting higher NK cell infiltration, greater micro-vessel density, and higher spontaneous apoptosis. Combining local NK cell delivery with therapy arrested tumor growth in the N1–S1 model. The machine learning models demonstrated high predictive accuracy, with R 2 values surpassing 0.90 in predicting key histopathological markers. The T1w and T2w models exhibited strong individual predictive capabilities, while the combined model demonstrated incremental benefits. The superior therapeutic response of the N1–S1 tumor model is attributed to a more immunogenic and vascularized TME, with the local administration of p NK cells being highly effective. This study validated the use of quantitative MRI and machine learning as powerful, non-invasive tools for predicting and monitoring TME changes, offering a promising framework for personalized HCC treatment and overcoming the limitations of conventional assessment methods.