Radiomics beyond the visible—Decoding tumor heterogeneity to guide combination therapy of intermediate-stage HCC

The integration of transarterial chemoembolization (TACE) with systemic immunotherapy represents a rapidly evolving paradigm in the management of intermediate-stage HCC, with possible applications across all disease stages, with dozens of trials underway.1 The recently completed EMERALD-1 and LEAP-012 trials2,3 have investigated the synergistic potential of combining TACE and systemic immunotherapy. While EMERALD-1 demonstrated modest benefits in progression-free survival (PFS) and some additive increments in tumor response rates for the experimental arm over the TACE plus placebo control arm, LEAP-012 failed to reach the PFS endpoint and is widely seen as a negative trial. Thus, strategies combining locoregional and systemic immunotherapy are not yet considered in guidelines such as the Barcelona Clinic Liver Cancer staging system,4 including its 2025 iteration. Initial subpopulation analysis of both trials shows promising signals in select cohorts, and the combination was clinically well tolerated. However, in the absence of reliable biomarkers, any potential benefits were not well reproducible. This is best explained by the profound biological and spatial heterogeneity of HCC. Decoding this heterogeneity and creating predictive biomarkers has therefore become key to transforming combination therapies from empirical successes into data-driven, personalized precision therapies. In this issue of Hepatology, Jin and colleagues5 present an ambitious, multicenter effort to bridge this gap. Their study introduces an imaging-based composite biomarker that quantifies both global and intra-tumoral heterogeneity (GTR–ITH index) to predict response and survival after combination therapy. This proof-of-principle pioneering work stands out due to its large scale, involving data from 742 patients across 19 tertiary centers, as well as methodological sophistication, linking advanced radiomics with immune-genomic phenotyping. THE CLOSER LOOK Jin and colleagues collected data from an impressive number of patients with primarily unresectable HCC undergoing combination therapy of TACE with systemic immunotherapy/molecular-targeted therapy. The authors designed a 2-part model capturing macroscopic tumor architecture (global tumor region, GTR) and microscopic heterogeneity (intra-tumoral heterogeneity, ITH). Both components are based on radiomic features extracted from pretreatment CT scans, computed either over the entire lesion (GTR) or across 100 lesion subdivisions to quantify spatial diversity (ITH). Using these features, the authors trained ensemble machine-learning models to predict treatment response. The resulting composite GTR–ITH index achieved consistent performance across internal validation and external testing datasets (AUC 0.82–0.83), demonstrating moderately strong correlation with objective response rates and overall and progression-free survival. The authors additionally explored the bio-pathological ground truth for their findings through transcriptomic correlation using TCGA-LIHC data. As such, the authors split the cohort into a low-risk (responding) group, which displayed immune-inflamed phenotypes characterized by enriched cytotoxic T-cell activity, M1 macrophage predominance, and higher immune/stromal scores, whereas high-risk (non-responding) tumors exhibited immuno-evasive characteristics with elevated M2 macrophages. The radio-genomic analysis captures an important and under-appreciated aspect in that it validates the clinical and imaging outcome findings and links survival outcomes and response to spatial immune profiling, which truly deserves the designation “biomarker.” Tumor heterogeneity is notorious in HCC and has long been recognized as a driver of treatment resistance and immune escape. Thus, the provided mechanistic insight addresses a clinically unmet need.6 CLINICAL IMPLICATIONS The translational implications of this proof-of-concept study have potential. In the best-case scenario, the GTR–ITH index could enable the pretreatment stratification of intermediate-stage HCC patients likely to benefit from combination therapies, guiding personalized treatment planning. As such, patients with a low GTR–ITH score, indicating a more homogeneous, immunologically “hot” tumor phenotype, may benefit from combination therapy. Conversely, those with high GTR–ITH scores can be considered likely to exhibit an immunologically “cold” phenotype that warrants alternative sequencing strategies (eg, neoadjuvant systemic therapy first or selective chemoembolization or thermal ablative boosting). The elegance of the proposed solution is the non-invasive nature of this biomarker, which makes it uniquely attractive and available both for initial monitoring and diagnosis as well as for serial and longitudinal monitoring of patients with evolving pathophysiology. In light of the rapidly evolving landscape of systemic therapies in HCC with numerous novel combinations being added to our armamentarium annually, serial quantification of intra-tumoral heterogeneity could reveal dynamic shifts under therapeutic pressure, potentially serving as an early surrogate for emerging resistance. This could be helpful for guiding systemic therapy choices between the various first-line and second-line combinations. An ambitious goal could even be to include this monitoring strategy in prospective trial designs. NOVELTY AND CHALLENGES What distinguishes this study from prior radiomic work is both its significant scale and its deliberate integration of imaging, machine learning, and immunobiology. The authors move beyond conventional texture analysis toward interpretable artificial intelligence, using Shapley additive explanations (SHAP)7 to identify feature contributions and radio-genomic validation to link imaging phenotypes to immune gene expression. This approach is well aligned with the global push toward fully interpretable, explainable predictive intelligence in cancer-directed therapy.8 Nonetheless, some important challenges and limitations remain. Ensemble learning models, while powerful, are complex and often opaque. Translating these signatures into clinically usable tools will require user-friendly platforms, regulatory standardization, and cross-vendor reproducibility of radiomic features, which remains a recurrent limitation in the field.9,10 The retrospective design of the study, while mitigated by external validation, introduces potential selection bias, and the pathological protocol bears the risk of heterogeneous, non-standardized assessment. In addition, the predominance of HBV-related HCC in the Chinese population may limit global generalizability, especially to Western cohorts where metabolic dysfunction–associated steatotic liver disease (MASLD) and HCV-related disease predominates. Further, although biologically correlated, the imaging–transcriptomic relationship remains associative rather than causal, and it is unclear whether the proposed biomarker is truly treatment-predictive or generally prognostic. Overall, the clinical adoption of such models will depend on integration within standard imaging workflows. Whether the GTR–ITH index can be implemented prospectively, in real time, and across varying scanner platforms remains an open question. CONCLUSIONS Jin and colleagues present a landmark contribution to the growing body of knowledge involving the interplay between imaging science and precision oncology. Their GTR–ITH index provides a quantitative, biologically grounded proof-of-principle framework for predicting outcomes in patients undergoing combination therapies of intermediate-stage HCC, establishing radiomics as a credible companion to genomic and immunologic biomarkers. While the translational potential remains unexplored, the study inaugurates a broader transformation in which imaging biomarkers, validated by pathological ground truth, have the potential to provide actionable guidance for molecular-targeted therapies. As combination therapies have the potential to reshape the HCC treatment landscape, the ability to noninvasively decode tumor heterogeneity will remain indispensable. The challenge ahead lies in translating such intelligent study designs into actionable clinical decision support systems. Jin and colleagues bring us a notable step closer to that goal.