Current interventional treatments for HCC comprise several locoregional therapies, each with unique mechanisms. TACE combines the targeted delivery of chemotherapy with the occlusion of arterial blood supply, thereby inducing ischemic necrosis in tumors by starving them of nutrients and oxygen, while ensuring a high concentration of cytotoxic agents [17]. Liver cancer ablation therapy encompasses thermal ablation techniques, including RFA and MWA. This procedure entails percutaneous access to intrahepatic lesions, employing either high or low temperatures to induce thermal or cryogenic coagulation and effectively destroying the tumor cells [17, 18]. Moreover, TARE delivers Yttrium-90 microspheres directly into the hepatic artery, providing localized radiation that damages tumor DNA while preserving most of the surrounding healthy tissue [19]. Together, these approaches form an integrated framework for the effective treatment of HCC (Fig. 1).

Fig. 1.

Schematic overview of interventional therapies in liver cancer.

Recent developments in interventional oncology have focused on enhancing the efficacy and safety of established therapies, as well as the introduction of innovative techniques such as irreversible electroporation (IRE) and nanoparticle-assisted therapies [2, 20]. IRE is a non-thermal ablation technique that disrupts cellular membranes using short, high-voltage pulses which lead to apoptosis [21]. The COLDFIRE-2 trial demonstrated that IRE is an effective and relatively safe salvage treatment for unresectable colorectal liver metastases (CRLMs) $\le$5 cm near critical structures, achieving 68% 1-year local tumor progression-free survival and 74% eventual local control, albeit with a 40% complication rate [22]. Nanoparticle-assisted therapies aim to improve the delivery and efficacy of chemotherapeutic agents and radio-sensitizers by leveraging the unique properties of nanotechnology for targeted cancer treatment [23]. For example, combining lyso-thermosensitive liposomal doxorubicin (LTLD) with focused ultrasound-induced hyperthermia safely enhanced intratumoral drug delivery by 3.7-fold in liver tumors, achieving targeted chemo-ablative effects without major toxicity [24].

Interventional therapies have demonstrated significant improvements in local tumor control and overall survival rates for select patient populations with HCC [25]. However, challenges persist in optimizing treatment selection, managing post-procedural complications, and addressing tumor recurrence. The heterogeneity of tumors, variability in patient response, and the evolving landscape of combination therapies necessitate continued research and the refinement of interventional strategies to improve clinical outcomes. The outcomes of interventional therapies can vary significantly due to factors such as tumor size and location, and the presence of underlying liver disease. Challenges in achieving complete tumor eradication and managing post-treatment recurrences are considerable obstacles in clinical practice [26].

The cellular constituents of the HCC TME include immune cells, endothelial cells, cancer-associated fibroblasts (CAFs), and other stromal components. These cells engage in dynamic crosstalk with the cancer cells, mediated by a plethora of signaling molecules such as cytokines, growth factors, and extracellular matrix (ECM) components [28]. The TME can foster a pro-tumorigenic environment, thus facilitating cancer cell proliferation, survival, and invasion.

Interventional therapies for cancer, such as chemotherapy, radiotherapy, and targeted therapies, have profound effects on the TME, which in turn affects therapeutic efficacy and tumor resistance (Fig. 2) [43]. A better understanding of these interactions is crucial for optimizing treatment strategies.

Fig. 2.

Complex interplay in the liver cancer tumor microenvironment (TME). This figure provides a comprehensive portrayal of the hepatocellular carcinoma (HCC) TME within a liver lobule, highlighting the cellular heterogeneity, immune cell diversity with both anti- and pro-tumor roles, the transition of fibroblasts to cancer-associated fibroblasts (CAFs), and tumor angiogenesis. The non-cellular matrix contains extracellular matrix (ECM) components and signaling molecule gradients. The effects of interventional therapies such as TACE and RFA/MWA on the TME are depicted schematically, showing altered signaling and cell fate. Finally, therapeutic modulation strategies such as immune checkpoint blockade and stromal disruption are represented to show TME “normalization” post-treatment. This figure encapsulates the dynamic and complex interactions occurring within the HCC TME, and which are essential for understanding and developing targeted therapeutic interventions.

The utility of AI in oncology is multi-faceted (Fig. 3), ranging from improved precision of diagnostic imaging to the personalization of therapeutic regimens, as well as the identification of novel targets for drug development.

Fig. 3.

Artificial intelligence (AI) and oncology interface. Upper panel: schematic representation of AI algorithms (e.g., neural networks, decision trees, support vector machines (SVM)). Middle panel: flowchart illustrating the process of data input (e.g., radiographic images, genomic profiles) into AI systems and output (e.g., diagnostic, prognostic, and therapeutic recommendations). Lower panel: examples of AI applications, such as radiomics, pathology image analysis, and multi-omics data integration.

In the realm of diagnostics, ML approaches can be broadly categorized into supervised and unsupervised learning paradigms [58]. Supervised models include Random Forests (RF), Gradient Boosting Machines (GBM) and Support Vector Machines (SVM). These are able to predict survival, classify tumors, and assess treatment response. Unsupervised models such as k-means and hierarchical clustering can be used to classify TME subtypes. Principal Component Analysis (PCA) is used to reduce variability in genomic and proteomic datasets. However, challenges exist due to class imbalances in clinical datasets, along with the need for strong cross-validation to avoid overfitting. Recently, the DL model has emerged as the predominant approach for diagnostic imaging. Convolutional neural networks (CNNs) that use hierarchical learning to detect subtle tumor features such as microvascular invasion or necrosis have demonstrated exceptional proficiency in interpreting radiographic images, sometimes exceeding the performance of experienced radiologists [59] and reporting higher accuracy for the detection of HCC [60]. EmerGNN is a flow-based graph neural network (GNN) approach that builds on conventional methodologies for predicting drug–drug interactions in emerging drugs by effectively leveraging biomedical networks [61]. Additionally, Vision transformers (ViTs) provided better classification results over histopathology slides by utilizing global attention to capture long-range dependences [62]. For the longitudinal imaging of liver tumors, recurrent neural network (RNN) models such as long short-term memories (LSTM) can manage temporal change [63]. As for reinforcement learning (RL), the RL-assisted design of reactive oxygen species-responsive polymeric nanocarriers was shown to enhance the anti-tumor efficacy of radiotherapy in HCC treatment [64].

Radiomics is the process of extracting many quantitative features from medical images (e.g., Computed Tomography (CT), MRI, PET) for data analysis and diagnosis support. By analyzing these features, clinicians are able to identify characteristic tumor patterns that could potentially be used to predict disease prognosis and treatment response, and to assess other clinically relevant outcomes. Radiomics transforms imaging data into quantifiable information, thus aiding personalized medical decision-making. For example, researchers have developed useful 3D-ResUNet models that enhance volumetric CT/MRI analysis [65]. Attention models assist with locating features from the heterogeneity of tumors. However, challenges persist due to datasets with limited annotation, and to the costs associated with computation of high resolution 3D models. Radiomics workflows typically involve a domain for disentangling, extracting, and quantifying certain features (e.g., texture, shape, wavelet) from imaging data using software tools, and then leveraging these quantified imaging features through predictive analytics with ML/DL models. Dynamic contrast-enhanced (DCE) MRI data can be used with temporal CNNs to longitudinally evaluate perfusion changes pre- and post-TACE treatment [66]. Standard pre-processing steps include normalization, artifact remediation, and sampling strategies for patch sampling of images to reduce the processing burden of large amounts of imaging data. However, the lack of standardization between institutions for radiomics feature extraction has led to issues with reproducibility.

The utility of AI also extends to the prognostic domain, where it can assimilate and analyze diverse datasets, including genomic, transcriptomic, proteomic, and clinical data. This integration of multi-omics and clinical information facilitates the development of predictive models that stratify patients according to their risk and likely response to treatment, thus guiding the selection of tailored therapeutic strategies [67]. Multi-omics approaches may also use GNNs to construct models for the multi-omic interactions (genes, proteins, and immune cells) occurring in the TME [68]. Multi-model transformers help to integrate imaging, genomics, and clinical data into clinical stratification. General challenges with data-fusion include the handling of batch effects across varying omics platforms, and having sufficient real scale and harmonized datasets to train robust models. AI-based interventional tools allow real-time processing, thereby enabling precision therapy. Although these AI methods demonstrate strong technical capabilities in the research setting, the next stage in their development is testing with real-world case examples.

In the therapeutic domain, AI applications are redefining drug discovery and development. AI platforms can rapidly screen vast chemical libraries to identify potential anti-cancer compounds, predict their efficacy, and optimize drug design, thus significantly accelerating the pipeline from laboratory bench to patient bedside [69].

Despite these promising advances, the integration of AI into clinical oncology is not without challenges. Ethical considerations such as patient privacy and data security, and the potential for algorithmic biases, necessitate careful scrutiny [70]. Furthermore, bridging the gap between the capabilities of AI and its practical implementation requires robust validation studies, regulatory oversight, and the education of healthcare professionals to work synergistically with AI tools [71]. As the potential of AI is harnessed, the associated ethical and logistical challenges must be navigated with prudence and foresight to ensure this technological revolution benefits patients across the spectrum of cancer care.

In order to integrate AI into interventional therapy for liver cancer, several data quality and heterogeneity challenges must first be addressed. Imaging protocols, biopsy sampling, and electronic health records can all introduce biases due to variable data collection protocols across institutions, thereby complicating the training of AI models. Additional problems are the lack of annotated datasets and class imbalance issues. With supervised learning, there is often a scarcity of correctly labeled data regarding health conditions, post-treatments, and HCC subtypes. Interpretability is another challenge, since “black box” DL algorithms such as CNNs can provide high accuracy in specific AI tasks, but may reduce clinician trust in AI models. Although regulatory bodies require both AI transparency and explainability, current interpretability methods may oversimplify the complicated biological interactions found in the TME, while requiring more computational resources. Moreover, the computational resource requirements can be challenging. In particular, the high-performance graphic processing units (GPUs) and cloud-based systems required by the 3D-ResUNet model to perform tumor segmentation are not feasible in many clinical settings. Data privacy issues are another consideration in multi-institutional collaborations. Regulations such as the General Data Protection Regulation (GDPR) and the Health Insurance Portability and Accountability Act (HIPAA) prevent the sharing of patient data because it would improve AI model generalizability. Population-specific biases embedded in trained AI models should also be considered. Moreover, TME systems evolve during treatment. This necessitates the use of adaptive AI models that learn continuously, which is not a feature of most current static models. Addressing these challenges will require standardization of data collection, the use of hybrid-interpretable AI, and optimization of hardware so that AI tools are sufficiently reliable for deployment in clinical practice in a transparent and efficient manner.

AI has transformed the diagnosis of HCC through unique image analysis. The diagnostic accuracy of AI algorithms is now on par with or even exceeds that of assessing radiologists across all major imaging modalities (Table 1, Ref. [72, 73, 74, 75, 76, 77, 78]).

CNNs analyze medical images through multiple layers. Early layers detect basic features like edges, while deeper layers identify complex tumor patterns. For example, studies have reported AUCs ranging from 0.89 to 0.92 with CNN models that assist in detecting HCC with CT-based diagnoses [72]. 3D-ResUNet enables volumetric segmentation through skip connections. These connections maintain spatial details while processing the full 3D context. This model yielded a Dice score of 0.88, with threshold tumors greater than 20 mm [73]. Radiomics complements DL by mathematically quantifying thousands of subvisual texture features from medical images. Radiomics-based analysis also improved HCC detection by predicting microvascular invasion with an AUC of 0.974 [74]. Multiparametric MRI images combined with DL technology achieved an extremely high AUC of 0.99 for lesion classification [75]. The addition of methods to assess peritumoral radiomics with a C-index of 0.73 to 0.77 improves recurrence prediction. Advanced ultrasound methods facilitated by AI have also been empowered. For example, contrast-enhanced ultrasound (CEUS)-based AI predicted multi-classification of focal liver lesions with an accuracy ranging from 79–92% [76], while highlighted real-time CNN visual predictive (analyses) for intra-procedural tumor targeting were associated with a 35% improvement in coverage. In short, the increasing incorporation of AI into all forms of medical imaging indicates ongoing transformation and improvement in the accuracy and precision of HCC diagnosis across imaging platforms.

AI-based radiomics is also changing the prediction of treatment response to TACE for HCC patients. Radiomics models use CT or MRI images to analyze features like texture, shape, and enhancement patterns. When combined with ML or DL, these models can predict which patients will benefit most from TACE. Several authors have demonstrated strong predictive performance for these AI models. For example, a segmentation model based on AI showed good performance, with average Dice scores of 83.05% for tumors and 92.72% for liver parenchyma. A combined clinical-radiomics model (using international normalized ratio (INR), albumin, and 10 radiomic features) performed better than clinical or radiomics models alone. It had an AUC of 0.878 for predicting post-TACE liver failure, compared to 0.785 (clinical model) and 0.815 (radiomics model) [77]. The clinical use of AI models has also demonstrated positive outcomes. For example, a 30% reduction in overtreatment was achieved by identifying non-responders early, and the use of real-time radiomics during TACE procedures led to a 35% improved coverage of tumors and 22% decrease in tumor recurrence [78]. These results indicate that durable and meaningful changes can be achieved by incorporating AI-based radiomics in clinical TACE treatment strategies for HCC patients, including prediction of tailored treatment response, and optimized treatment planning.

Collectively, these studies underscore the transformative impact of AI when integrated with interventional therapy. By harnessing the power of AI to interpret vast and complex data from HCC, clinicians can make informed decisions that lead to more precise and effective treatment. As AI algorithms continue to be refined and integrated into the clinical setting, we anticipate a new era of personalized and adaptive oncology care that is driven by deeper insights into the biology of the TME.

The TME is a complex mosaic of cellular and acellular components that play pivotal roles in tumorigenesis, metastasis, and response to therapy. Deciphering the interplay between different TME constituents within this intricate milieu and identifying potential therapeutic targets are critical for improving cancer treatment. The application of AI has emerged as an invaluable approach in untangling the complexities of the TME and identifying novel targets for interventional therapies (Table 2, Ref. [76, 77, 78, 79, 80]).

AI, and especially DL and ML, has reshaped the prediction of immunotherapy response in HCC, enabling the analysis of complex features of the TME and biomarker signatures. For example, Zhang et al. [76] created a DL model that analyzed histopathology images of HCC to identify immune-associated gene signatures (PD-L1, CD8A) that correlated with response to atezolizumab/bevacizumab combination therapy, achieving an AUC of 0.92. Liang et al. [77] produced an interpretable DL framework called Pathological-biomarker-finder that identified novel tissue biomarker signatures associated with immune activity in the TME with 88% accuracy. Another pivotal development is the quantification of CD8 TILs using AI. The ATLS-8 method stratifies patients into prognostic groups. Those with high ATLS-8 had a 5-year overall survival (OS) of 65.03%, while those with low ATLS-8 had a 5-year OS of 34% (HR = 2.40, 95% CI: 1.37–4.19, p = 0.015) [79]. AI has also improved radiomics-based models, with clinical-radiomics models performing better than solely clinical-based models (AUC: 0.89 vs. 0.60, p 0.05). Certain features, such as peritumoral arterial enhancement and tumor distribution, were significant predictors of treatment response.

AI has the capacity to integrate and interpret vast amounts of data derived from the TME, including but not limited to genomic, transcriptomic, proteomic, and imaging data. Such integrative analysis can reveal hidden patterns and biomarkers that are indicative of tumor behavior and therapeutic response. Multi-omics integration has also been instrumental in transition, as demonstrated by Xie et al. [78] who employed AI to integrate genomics and radiomics data. These authors successfully distinguished immune-excluded (‘cold’) tumors from immune-inflamed (‘hot’) tumors, which are classifications associated with the response to PD-1 inhibitors. Dai et al. [80] demonstrated the accuracy of AI integration into prognostic formatting by employing LASSO regression to create an immune-related gene prognostic index (IRGPI). This index predicted TME immune cell infiltration and the efficacy of immunotherapy with a C-index of 0.75. By applying interferon-gamma and antigen presentation gene expression signatures, DL models can stratify HCC patients into either immunotherapy responders or non-responders (e.g., nivolumab clinical trials) [81].

Beyond predicting the response to immune-based therapy, AI has also played a role in the prediction of treatment effectiveness. Combining radiomic features with clinical data in CT- and MRI-based DL models were found to improve the prediction of response to therapy. Nevertheless, significant challenges remain, including data heterogeneity, the ‘black-box’ nature of CNNs, and prospective validation. It will also be critical to standardize datasets and develop explainable AI (XAI) models to promote their clinical application [82]. Overall, AI has demonstrated multi-faceted roles in elucidating the complexity of the TME. It offers a lens through which the subtle and intricate interactions occurring within the TME can be observed and interpreted. By leveraging AI to reveal the underlying biological processes and to identify novel therapeutic targets, the precision and effectiveness of interventional therapies for liver cancer can be improved.

The notion of manipulating the TME to enhance the efficacy of cancer therapies is an area of active investigation. AI-driven approaches are increasingly being employed to model the TME and predict how manipulations might influence treatment outcomes. Such approaches include the use of computational simulations to predict the impact of altering TME factors concurrent with interventional strategies.

AI algorithms have the potential to simulate a variety of TME manipulations, ranging from the targeting of specific cell populations to the modulation of signaling pathways. For example, Mpekris et al. [83] utilized AI to model the vascular structure of the TME and its response to anti-angiogenic therapy in combination with chemotherapy. This computational approach helped to predict how changes in the vascular landscape of the TME could improve drug delivery and efficacy. Another AI-driven method employs agent-based modeling to simulate interactions within the TME, including the response of immune cells to immunotherapies. For example, an AI model was used to simulate the effects of immune checkpoint inhibitors on the spatial and temporal dynamics of T cells within the TME [84]. This predictive model provided insights into how reinvigoration of the immune system could be optimized when used in combination with other interventions.

AI has also been instrumental in predicting how genetic and epigenetic modification of the TME can influence therapy response. AI models were used to identify epigenetic alterations in the TME that could serve as biomarkers for response to epigenetic therapy [85]. By integrating multi-omics data, AI can guide the development of therapeutic strategies that modulate TME to improve patient outcomes.

In conclusion, AI-driven approaches are proving invaluable for modeling TME manipulation and providing a predictive framework for concurrent interventions. By simulating the dynamic and complex interactions within the TME, AI can help to design more effective combination therapies, ultimately paving the way for personalized treatment plans that are responsive to the unique characteristics of each patient’s TME.

AI is likely to revolutionize the personalization of interventional treatments in oncology over the coming years. It is poised to play a central role in tailoring therapies to individual patient profiles by taking into account the unique characteristics of their TME. Here, we explore some of the novel applications for AI in the crafting of interventional treatments.

One of the most promising directions for AI is the integration of multi-omics data to inform treatment decisions. The seminal work of Chen and Snyder [86] highlighted the potential use of omics data to obtain a holistic view of the patient’s biology. AI can synthesize information from genomics, transcriptomics, proteomics, and metabolomics to construct comprehensive models of the TME. This multi-faceted approach enables the identification of patient-specific molecular signatures that can predict response to specific interventional therapies.

The expanding field of spatial multi-omics is another area where AI is set to make a significant contribution. By combining high-dimensional data with spatial information, AI can map the complex organization of the TME and identify loci that are amenable to targeted intervention. A study on the use of spatial transcriptomics combined with AI analysis to map the cellular heterogeneity of tumors was described recently, providing insights into the spatial distribution of therapeutic targets [87].

In the drug development field, AI can accelerate the discovery of novel agents that modulate the TME to improve the efficacy of interventional treatments. Stokes et al. [88] reported on the use of DL to identify a new antibiotic compound through the analysis of bacterial gene expression profiles. A similar approach could be employed to discover agents that are able to reshape the TME, thereby increasing the effectiveness of interventional oncology treatments.

AI algorithms are also being developed to predict and monitor evolution of the TME in real-time during treatment, allowing for dynamic adjustments to therapy. The predictive modeling capabilities of AI can anticipate changes in the TME in response to treatment, enabling clinicians to modify the intervention strategy in order to avoid resistance or to enhance therapeutic efficacy.

In future, AI will become an indispensable tool in the personalization of interventional treatments. By harnessing the power of AI to interpret the wealth of biological data available, a new paradigm for cancer treatment arises whereby patients receive therapy that is precisely tailored to their individual TME, leading to improved outcomes and a higher quality of life.