Nuclear medicine in predicting hepatocellular carcinoma response

Abstract

Immune checkpoint inhibitors and anti-angiogenic targeted therapies have improved outcomes in hepatocellular carcinoma (HCC), but responses remain heterogeneous, creating a need for noninvasive biomarkers to enable early treatment adaptation. We review clinical and translational evidence on nuclear medicine approaches – PET/computed tomography (CT), single-photon emission computed tomography (SPECT), radiomics, machine learning, and theranostics – for response prediction and prognostication in HCC treated with immunotherapy alone or in combination with targeted agents. Metabolic PET/CT, most commonly with ¹⁸F-fluorodeoxyglucose, supports pragmatic risk stratification; volumetric indices such as metabolic tumor volume (MTV) and total lesion glycolysis generally provide stronger prognostic enrichment than single-voxel metrics, and an MTV threshold of greater than or equal to39.65 cm³ has been reported to associate with poorer outcomes. Immune-targeted PET/SPECT extends beyond metabolism by mapping target availability and heterogeneity (e.g. PD-L1) and immune activation or effector function (e.g. CD137, granzyme B), although current studies are often small and retrospective. PET-based radiomics and machine learning can generate imaging surrogates of immune phenotypes and aggressive biology, but reproducibility is limited by acquisition/reconstruction differences, segmentation variability, and scarce external validation. Theranostics offers an image-guided ‘select-and-treat’ paradigm for radionuclide therapy, yet target heterogeneity, dosimetry standardization, cost, and infrastructure remain barriers. Translation to routine care will require harmonized protocols, multicenter prospective validation, and demonstration of decision impact.

Introduction

Hepatocellular carcinoma (HCC) remains a leading cause of cancer-related mortality worldwide, with a persistently high global burden. Epidemiologic estimates from the IARC Global Cancer Observatory highlight liver cancer as one of the major contributors to cancer deaths, underscoring the ongoing need for effective systemic strategies and optimized patient selection [1]. Over the past few years, first-line systemic therapy has shifted toward combination regimens that integrate immunotherapy and targeted agents. In the IMbrave150 trial, atezolizumab plus bevacizumab improved survival outcomes compared with sorafenib, establishing an immunotherapy-based standard for unresectable HCC [2]. Likewise, the HIMALAYA study demonstrated survival benefit with a tremelimumab–durvalumab regimen, further expanding immune-based options in the frontline setting [3].

Despite these advances, clinical benefit remains heterogeneous: a substantial proportion of patients experience primary resistance or early progression, while others derive durable responses. This variability motivates the search for predictive biomarkers that can be applied before or early during treatment to guide regimen choice and avoid ineffective therapy. However, conventional tissue-based biomarkers (e.g. PD-L1 immunohistochemistry) face practical and biological constraints in HCC, including limited tissue availability, sampling bias, and pronounced intra- and inter-lesional heterogeneity. These limitations are particularly relevant when clinical decisions are made for multifocal disease or in patients not routinely undergoing repeat biopsies.

Nuclear medicine offers a complementary paradigm by enabling whole-body, quantitative, and repeatable characterization of tumor biology and the tumor microenvironment. At the metabolic level, ¹⁸F-fluorodeoxyglucose (¹⁸F-FDG) PET/computed tomography (CT) has been widely explored for prognosis and early risk stratification, yet its specificity for immune response remains imperfect and may be confounded by tumor differentiation, inflammation, and background liver disease. Meanwhile, emerging immuno-PET strategies can noninvasively map target availability and spatial heterogeneity across all lesions; for example, PD-L1 immuno-PET with ⁸⁹Zr-labeled therapeutic antibodies has shown that tracer uptake across lesions can be associated with clinical outcome under PD-(L)1 blockade, supporting the concept of ‘imaging-informed’ patient enrichment. In parallel, consensus efforts in nuclear medicine have emphasized the need to standardize response assessment frameworks (beyond RECIST: Response Evaluation Criteria in Solid Tumors) and to interpret metabolic changes in the context of immunotherapy-specific patterns. Figure 1 summarizes the clinical rationale for incorporating nuclear medicine as a screening and stratification tool: by identifying biological phenotypes associated with benefit vs. non-benefit, imaging may support more personalized selection among immunotherapy-based regimens and alternative therapies.

Fig. 1.

Nuclear medicine-guided stratification for precision systemic therapy in hepatocellular carcinoma (HCC). Schematic overview of how molecular imaging can support patient-level and lesion-level stratification in HCC treated with immunotherapy-based regimens. In the absence of biomarker-guided selection, clinical benefit is heterogeneous. PET/SPECT readouts of tumor biology and the tumor microenvironment (TME) provide whole-body information that may help identify patients more likely to benefit from immunotherapy/targeted combinations vs. those who may require alternative strategies or early treatment adaptation. SPECT, single-photon emission computed tomography.

PD-L1 immuno-PET for PD-(L)1 blockade: whole-body target accessibility, heterogeneity, and clinical validity constraints

Figure 2 outlines the general workflow of radiolabeled antibody-based imaging, from tracer administration and target engagement to PET-based evaluation, with potential extension to radionuclide therapy in a theranostic framework. PD-L1 immuno-PET was developed to address a practical limitation of tissue PD-L1 immunohistochemistry: a single biopsy provides an inherently spatially restricted and temporally static measurement that can be discordant with whole-body disease biology, especially when PD-L1 expression is heterogeneous across lesions or dynamically regulated by therapy. By contrast, antibody-based PD-L1 PET aims to quantify whole-body target accessibility and lesion-to-lesion heterogeneity, potentially providing a more clinically relevant representation of whether PD-L1-directed therapy can engage its target across the entire tumor burden. The strongest clinical rationale for this approach comes from outcome-linked exploratory imaging with ⁸⁹Zr-atezolizumab in a small translational cohort (n = 22), in which whole-body uptake metrics [including geometric aggregation of lesion standardized uptake value (SUV)] were associated with subsequent outcomes under PD-L1 blockade and a tentative SUV-based stratification scheme (e.g. a geometric mean SUVmax cutoff) was proposed to separate patients with more favorable trajectories from those less likely to benefit [4]. While this study is frequently cited as ‘predictive’, a rigorous interpretation is that it provides hypothesis-generating clinical validity, because its small sample size, mixed tumor biology, and workflow sensitivity (protein mass dose, imaging time window, reconstruction, and physiologic lymphoid uptake) mean that any proposed cutoff should not be treated as transferable without external prospective validation [4].

Fig. 2.

General workflow of immuno-PET and theranostic extension using radiolabeled antibodies or antibody-derived agents. After intravenous administration, radiolabeled targeting vectors distribute systemically and bind immune-related targets (e.g. PD-1/PD-L1) in tumors and lymphoid tissues. PET imaging enables noninvasive assessment of target availability and spatial heterogeneity across lesions, which can inform response assessment and hypothesis-generating biomarker development. In selected settings, the same or a matched targeting vector can be labeled with a therapeutic radionuclide (e.g. β-emitter) to enable image-guided radionuclide therapy and post-therapy biodistribution/dosimetry evaluation, constituting a theranostic strategy.

A major reason that PD-L1 immuno-PET thresholds are difficult to standardize is that PD-L1 is not confined to tumor cells: normal lymphoid tissues (spleen/lymph nodes) and other immune compartments can show high physiologic uptake, creating a ‘sink effect’ that compresses tumor dynamic range and makes quantitative readouts strongly dependent on mass dose and timing. This was mechanistically clarified in preclinical work using ⁸⁹Zr-DFO-avelumab, which demonstrated high uptake in spleen and lymph nodes and showed a marked mass effect with cold antibody co-injection: competitive blocking reduced splenic binding while altering nodal and tumor contrast in a dose-dependent manner, emphasizing that SUV/TBR (tumor-to-background ratio) cannot be compared across studies unless protein dose, imaging schedule, and background correction are explicitly controlled [5]. Complementary preclinical data with ⁸⁹Zr-DFO-REGN3504 reinforced that immuno-PET intrinsically images both tumor and immune compartments and that splenic uptake is not merely ‘noise’: macrophage depletion experiments reduced splenic uptake (e.g. from 29.3 to 8.9%ID/g, P = 0.0012), supporting an immune-cell/antibody-handling contribution to background signal and highlighting why quantification must account for immune organ biodistribution [6]. Together, these studies justify a critical review stance: high lymphoid uptake is biologically meaningful but methodologically challenging, and it is a primary source of heterogeneity that limits cross-center reproducibility unless standardized protocols and modeling strategies are adopted [5,6].

Engineering and labeling innovations have been proposed to expand the usable dynamic range and improve interpretability. The Probody concept (tumor-restricted activation) was designed to reduce peripheral PD-L1 engagement and thereby improve tumor-to-background contrast. In preclinical comparisons, ⁸⁹Zr-CX-072 showed improved tumor enrichment relative to conventional PD-L1 antibodies, including a reported improvement in tumor-to-blood metrics (e.g. a 2.1-fold increase in tumor-to-blood ratio at later timepoints) and substantially lower levels of activated form in spleen relative to tumor (reported as several-fold differences), supporting its potential to mitigate lymphoid sink effects [7]. Importantly, this ‘solution’ introduces a new biological dependency – activation requires tumor protease activity – which is itself heterogeneous across tumor types, lesions, and treatment states; therefore, improved contrast does not automatically imply improved prediction, and the possibility of context-dependent false negatives must be considered when translating Probody tracers to diverse clinical cohorts [7]. In parallel, site-specific labeling and dosing strategies have aimed to stabilize quantitative performance. With ⁸⁹Zr-DFO-6E11, cold antibody co-injection markedly reduced splenic uptake (e.g. 14.44 ± 3.10 to 5.59 ± 1.52%ID/g, P < 0.001) in xenograft models, illustrating that background is partially controllable and reinforcing that mass dose should be treated as a prespecified parameter rather than a hidden variable [8]. However, such approaches also raise translational caveats: changing mass dose and target occupancy may alter pharmacokinetics and potentially perturb immune biology, particularly in the context of concomitant immunotherapy, so ‘background optimization’ must be evaluated not only for imaging contrast but also for biological neutrality and clinical practicality [8].

Early clinical feasibility studies further illustrate that the principal near-term utility of PD-L1 immuno-PET is establishing interpretable biodistribution and informing trial design, rather than delivering validated clinical decision thresholds. In the first-in-human evaluation of ⁸⁹Zr-CX-072, imaging characterized time-dependent tumor conspicuity and background behavior under different protein mass conditions, supporting feasibility and offering a protocol foundation for subsequent outcome-anchored studies, but not providing robust predictive metrics due to limited sample size and heterogeneous tumor types [9]. Similarly, microdose exploratory imaging with ⁸⁹Zr-DFO-KN035 (single-domain PD-L1 binder) in a small clinical cohort (n = 12) proposed that PD-L1 immuno-PET might support not only tumor stratification but also mapping of tissue patterns potentially associated with prognosis and immune-related adverse events; nevertheless, the single-arm design, small event counts, and lack of prespecified, standardized endpoints constrain inference and should be framed as exploratory signal generation rather than confirmed prediction [10]. Translational bridging studies in patient-derived platforms add plausibility without fully resolving clinical validity. In RCC tumorgraft models spanning wide PD-L1 immunohistochemistry expression (0–85%), ⁸⁹Zr-atezolizumab uptake metrics correlated with immunohistochemistry in blinded analyses (reported correlations with statistical significance), while only a few paired patient illustrations were available, indicating that the work strengthens quantitative concordance but still falls short of outcome-validated clinical prediction [11].

Finally, mechanistic evidence cautions against oversimplified interpretation of PD-L1 imaging signals as purely ‘tumor-cell PD-L1’. A study integrating targeted PD-L1 degradation with PD-L1 PET suggested that host (non-tumor) PD-L1 may be a dominant determinant of anti-PD-1 efficacy in certain contexts, implying that immuno-PET uptake may integrate contributions from tumor cells and host immune/stromal compartments and thereby help explain discordance between single-biopsy immunohistochemistry, PET signal, and clinical outcomes [12]. This line of evidence is particularly useful in a critical review because it reframes ‘imaging–immunohistochemistry mismatch’ as potentially informative rather than simply erroneous, while also underscoring the need for compartment-aware validation strategies.

In summary, PD-L1 immuno-PET provides a conceptually strong approach to visualize whole-body target accessibility and heterogeneity, and early outcome-linked clinical evidence supports its potential role as a complementary biomarker to tissue immunohistochemistry [4]. However, across studies, clinical validity remains constrained by small cohorts, protocol sensitivity, physiologic lymphoid uptake and sink effects, and biological context dependencies introduced by tracer design (including Probody activation) [12]. The most defensible current conclusion for an science citation index-level review is that PD-L1 immuno-PET is clinically promising but not yet a standardized predictive tool: multicenter prospective trials with harmonized acquisition/reconstruction, prespecified mass doses and imaging windows, and predefined quantitative thresholds linked to clinically meaningful endpoints (ORR: objective response rate; PFS: progression-free survival; OS: overall survival; and irAE: immune-related adverse event mapping) are required before PD-L1 immuno-PET can move from exploratory association to actionable decision support [11].

PD-L1 single-photon emission computed tomography/computed tomography with ⁹⁹mTc-NM-01: a translational tracer bridging molecular specificity and clinical accessibility

⁹⁹mTc-NM-01 (PD-L1 SPECT/CT) is a translational tracer that extends PD-L1-targeted nuclear imaging from PET toward the more widely accessible SPECT/CT platform. The probe showed high affinity for human PD-L1 (Kd = 0.8 nM) and did not compete with atezolizumab, supporting potential compatibility with PD-(L)1 therapy settings. In a PD-L1-positive xenograft model (HCC827), SPECT/CT demonstrated visible tumor uptake with quantitative biodistribution of 1.50 ± 0.27%ID/g (30 min) and 1.23 ± 0.18%ID/g (90 min), predominantly renal clearance, and biphasic blood clearance (t½ 2.1 and 25.4 min). Specificity was supported by blocking experiments (excess unlabeled NM-01 reduced tumor contrast). Toxicology testing suggested a favorable translational profile, with a reported maximum tolerated dose of 2.58 mg/kg. Overall, the study provides a solid feasibility and development package for PD-L1 SPECT/CT, but clinical validity remains unproven: evidence is mainly preclinical, SPECT quantification is less robust than PET, and PD-L1 imaging is susceptible to physiologic lymphoid/inflammatory uptake – necessitating larger prospective, outcome-anchored clinical validation [13].

CD8⁺ T-cell immuno-PET: first-in-human ⁸⁹Zr-Df-IAB22M2C (feasibility, biodistribution, dosimetry)

CD8-targeted immuno-PET (⁸⁹Zr-Df-IAB22M2C) provides first-in-human evidence that whole-body imaging of CD8⁺ T-cell-rich tissues and tumor lesions is feasible. In this Phase I study (n = 15, heterogeneous solid tumors), 36/38 lesions demonstrated discernible tumor uptake. Lesion activity was highly variable (SUVmax 3.7–58.0), underscoring substantial inter-tumor and microenvironmental heterogeneity; notably, a HCC case showed SUVmax 19.3. Physiologic uptake was prominent in immune organs such as spleen and bone marrow, consistent with expected CD8 distribution but also indicating that interpretability depends on optimized imaging windows and tumor-to-background contrast. Dosimetry was ~0.64 mSv/MBq (dose-escalation) and 0.67 mSv/MBq (dose-expansion), highlighting the need to balance information gain against radiation burden. Critically, this work primarily establishes safety, biodistribution, and feasibility rather than predictive utility: the sample size is small with marked tumor-type heterogeneity, and systematic correlations with histologic CD8 density or immunotherapy outcomes (ORR/PFS/OS) were not established, so clinical validity as a response biomarker requires larger, tumor-specific prospective validation [14].

Immune-activation imaging: CD137 (4-1BB) PET for early immune checkpoint inhibitor response in hepatocellular carcinoma (preclinical + pilot clinical)

Cheng et al. developed [¹⁸F]AlF-NOTA-BCP137 as a CD137 (4-1BB) PET tracer to image early immune activation during immune checkpoint inhibition. The study combined preclinical validation with a pilot clinical cohort of five patients with HCC scanned at baseline and early during immune checkpoint inhibitor (ICI) therapy. Using an exploratory positivity threshold of tumor-to-muscle SUVmax ratio (T/M SUVmax) > 3, 2/5 patients were classified as CD137-positive and showed more favorable clinical outcomes, whereas the remaining CD137-negative patients had poorer outcomes. This work is valuable for a ‘molecular imaging-based immunotherapy prediction’ narrative because CD137 is a functional activation marker more proximal to effector T-cell engagement than static checkpoint expression. However, the clinical evidence level is necessarily limited: the sample size is extremely small (n = 5), the T/M > 3 cutoff is hypothesis-generating, and readouts may be sensitive to scan timing, liver background activity, prior treatments, and tumor burden; therefore, reproducibility and generalizability across centers and tumor contexts remain unproven and require larger prospective validation before clinical deployment [15].

Immune function imaging: granzyme B PET

Granzyme B (GzmB) PET shifts immunotherapy imaging from checkpoint expression to an effector function readout – that is whether cytotoxic immune killing is actually occurring. In a prospective translational animal study, ⁶⁸Ga-/GZP-based GzmB PET was used across multiple mouse tumor models to track therapy-induced GzmB activity and relate the PET signal to tumor control and immune infiltration, supporting its role as a mechanistically proximal biomarker of immune-mediated tumor killing rather than a static surrogate of target availability. Importantly, the authors restricted analyses to tumors above a minimum volume threshold (≈40 mm³) and did not apply multiple-comparison correction, reinforcing that the statistical signals should be interpreted as hypothesis-generating rather than a transferable clinical cutoff [16]. A related study in an HCC model (Hepa1-6) using [¹⁸F]AlF-mNOTA-GZP further suggested that GzmB PET can differentiate responders from nonresponders earlier than tumor volume changes, aligning with the concept that effector-function imaging may provide earlier response stratification than target-expression imaging [17]. Methodologically, recent patent-oriented work highlights cyclic peptide ligands targeting GzmB as a next-generation tracer direction; however, this evidence should be framed strictly as developmental potential rather than clinical proof, given the absence of systematic human performance data and the known challenges of GzmB imaging (time-window dependence, inflammation vs. durable killing, and clearance/background considerations) [18]. Overall, GzmB PET is a compelling ‘effector-end’ immuno-imaging strategy with strong mechanistic plausibility, but current evidence remains predominantly preclinical and protocol-sensitive, requiring standardized workflows and prospective human validation against ORR/PFS/OS before it can be considered a clinically established predictive biomarker [16–18].

Other immune checkpoints/activation targets: LAG-3 PET as an early readout of combination immunotherapy (preclinical hepatocellular carcinoma)

A 2025 preclinical study in murine HCC developed the LAG-3-targeted PET tracer [⁶⁸Ga]Ga-NOTA-C25 to noninvasively monitor LAG-3-associated signals in tumor-infiltrating immune cells during combination immunotherapy (e.g. STING agonism with anti-PD-1 alone or in combination with anti-LAG-3). The study provides a relatively complete ‘nuclear quantification–immunology–therapy effect’ chain: tracer uptake correlated strongly with LAG-3+ tumor-infiltrating lymphocytes (R² = 0.855) and dynamically reflected early immunomodulation. Quantitatively, tumor uptake (%ID/g) at day 4 differed substantially across groups (PBS 0.391 ± 0.061, STING 0.851 ± 0.037, and STING + αPD-1 1.60 ± 0.374), while by day 13 uptake increased in controls (PBS 0.647 ± 0.09) but decreased in effective combination therapy (e.g. STING + αPD-1 0.402 ± 0.017), consistent with evolving immune activation followed by suppression/exhaustion remodeling captured by LAG-3-targeted imaging. Critically, the evidence remains preclinical: %ID/g is not directly transferable to clinical SUV, tumor models and combination regimens may not represent the heterogeneity of human HCC, and whether LAG-3 PET predicts durable endpoints (e.g. OS) or generalizes across immunotherapy combinations requires multi-model validation and prospective clinical cohorts [19].

Fibroblast activation protein/cancer-associated fibroblast axis and fibroblast activation protein inhibitor imaging (stromal/tumor microenvironment)

In metastatic colorectal cancer (n = 131), ⁶⁸Ga-fibroblast activation protein inhibitor (FAPI) PET was used as a quantitative surrogate of the cancer-associated fibroblast (CAF)/fibroblast activation protein (FAP) stromal axis to monitor microenvironment remodeling during combined transforming growth factor-βR inhibition and immunotherapy. Imaging uptake metrics (e.g. SUV-based measures/TBR) were linked to tissue biology, showing a positive correlation with FAP protein expression (figure-annotated r ≈ 0.5455, P ≈ 0.0422, as labeled) and inverse associations with immune infiltration including CD8⁺ T-cell-related measures, supporting the concept that a stroma-rich (‘immune-cold’) phenotype can be captured noninvasively; posttreatment reductions in FAPI signal were reported to align with increased CD8 infiltration. Key caveats are that correlation strength depends on sampling timing/site and organ-specific background, and – when cited in an HCC-focused review – this work should be framed as mechanistic/methodological evidence rather than direct clinical decision guidance in HCC [20].

Complementing these findings, a prospective conference abstract in HCC enrolled 64 patients who underwent baseline [¹⁸F]AlF-NOTA-FAPI-04 PET/CT, with paired tumor/adjacent tissues available in 22 cases for immunohistochemical assessment of fibrosis and CAF-related markers. A subset of advanced-stage patients underwent an early re-scan after two cycles of targeted therapy plus immunotherapy to evaluate dynamic changes in tracer uptake. Using exploratory Kaplan–Meier and Cox regression analyses, the authors reported that PET-derived metrics including TSUVmean and Boolean-operation-derived tumor-to-liver fibrosis were independently associated with PFS (hazard ratios and P values were provided in the abstract). Here, Boolean-operation-derived tumor-to-liver fibrosis was described as a Boolean-operation-derived, liver fibrosis-related FAPI PET parameter computed from tumor region of interest: ROIs (ROI), total-liver ROIs (TROI), and background ROIs (BROI). Given that this evidence comes from a meeting abstract rather than a full peer-reviewed publication, it should be interpreted as preliminary and hypothesis-generating, with limited methodological transparency, and should not be used as the sole basis for definitive conclusions [21].

Metabolic tracer comparison: ¹¹C-acetate vs. ¹⁸F-fluorodeoxyglucose PET/computed tomography linked to immune infiltration and prognosis in hepatocellular carcinoma

In a dual-tracer cohort of HCC, 39 patients underwent both scans and 32 with surgical pathology/immune profiling were analyzed to connect metabolic phenotype-immune infiltration-prognosis. Quantitatively, median tumor uptake for ¹¹C-acetate was SUVmax 5.38 (interquartile range 3.46–7.52) with TBR 1.00 (0.74–1.30), whereas ¹⁸F-FDG showed SUVmax 2.62 (1.96–4.73) and TBR 0.95 (0.70–1.34). An immune score (Immunoscore) was associated with the ¹⁸F-FDG metabolic phenotype, and low Immunoscore predicted worse OS (P = 0.002); FDG-derived metrics also correlated with immune markers including CD3, CD8, and granzyme B, supporting biological coupling between glucose metabolism and immune-related features. Key limitations are the small effective sample size (n = 32), limited robustness for subgroup inference, and that endpoints were immune infiltration and survival rather than ICI response; additionally, dual-tracer workflows – especially ¹¹C-acetate – reduce scalability and may introduce selection bias [22].

Metabolic–immune phenotype links: ¹⁸F-fluorodeoxyglucose SUVmax as a correlate of PD-1/PD-L1 (association, not immune checkpoint inhibitor prediction)

Several retrospective clinical studies suggest that an FDG-high metabolic phenotype is statistically associated with PD-1/PD-L1 expression, but this evidence should be interpreted as imaging–immune correlation rather than validated prediction of ICI benefit. In bladder cancer (n = 63), preoperative ¹⁸F-FDG PET/CT showed significantly higher SUVmax in PD-1-positive and PD-L1-positive tumors, and SUVmax remained independently associated with PD-1/PD-L1 status on multivariable analysis; an exploratory cutoff of SUVmax = 22.7 was proposed to classify PD-1 and PD-L1 expression with accuracies of 71.4% and 77.8%, respectively [23]. In HCC (n = 102), tumor SUVmax correlated positively with PD-L1 immunohistochemistry (Spearman r = 0.421, P < 0.001), and the PD-L1-high group had substantially higher SUVmax (4.04 ± 1.17 vs. 7.85 ± 2.77), with additional associations reported between SUVmax and immune-microenvironment features (e.g. macrophage-related markers) and PD-L1 status in multivariable models [24]. Similarly, in colorectal cancer liver metastases after curative colon surgery (n = 72), SUVmax correlated with PD-L1 (r = 0.458, P < 0.001) and receiver operating characteristic (ROC) analysis suggested moderate discrimination (AUC: area under receiver operating characteristic was 0.745) with an optimal cutoff of SUVmax 6.15 (sensitivity 65.2%, specificity 75.5%) for PD-L1 classification [25]. Critically, across these studies the endpoint is PD-1/PD-L1 expression, not ICI outcomes (ORR/PFS/OS); therefore, SUVmax thresholds should be framed only as potential enrichment/screening signals that are vulnerable to biopsy/immunohistochemistry scoring variability, intra-/inter-lesion heterogeneity, and imaging-protocol dependence, and they require prospective validation in ICI-treated cohorts before being considered clinically actionable predictors [24].

Fluorodeoxyglucose-PET for prognosis/early failure risk stratification: from intensity to metabolic burden (metabolic tumor volume/total lesion glycolysis)

FDG-PET is not an immune-target-specific modality, yet pretreatment FDG-derived metrics repeatedly show clinical utility for baseline risk stratification in patients receiving ICI-based regimens. A consistent signal across studies is that volumetric indices reflecting global metabolic tumor burden [e.g. metabolic tumor volume (MTV) and total lesion glycolysis (TLG)] tend to outperform single-voxel intensity measures (SUVmax) for prognostic separation, likely because they better integrate overall disease load and biological aggressiveness. However, the current evidence base is dominated by single-center retrospective cohorts, with cutoffs frequently derived by within-cohort ROC optimization; accordingly, reported thresholds should be interpreted as cohort- and protocol-dependent and require external validation before being used for treatment decisions [26–31]. Key imaging metrics, thresholds, and clinical endpoints are summarized in Table 1.

Table 1.

FDG-PET for prognosis/response prediction: volumetric burden (MTV/TLG) and related metrics

Study (cancer type)	Design/N	Treatment/endpoint	Key PET metric(s) and cutoff (if reported)	Main finding (clinical meaning)	Critical limitations (why it’s not definitive)
Evangelista et al. (non-small cell lung cancer) [26]	Retrospective/32	Nivolumab; RECIST clinical benefit vs. PD	SUVmaxwb + LDH; favorable stratum: SUVmaxwb < 13.5 and LDH < 200 U/l	Low baseline whole-body metabolic activity + low LDH enriched for clinical benefit (simple, ‘low-barrier’ stratification).	Small single-center; cutoffs likely cohort-dependent; RECIST may misclassify pseudo-progression.
Wang et al. (uHCC, conversion) [27]	Retrospective; recruited 62, analyzed 20	PD-1 inhibitor + lenvatinib; pathological response	Higher SUVmean/SUVpeak/TLR in poor responders; tumor size and TLR independent predictors	Links baseline FDG phenotype to hard endpoint (pathology) after ICI + TKI conversion therapy.	Extreme selection bias (62 → 20); regimen-specific; thresholds/protocol reproducibility not established.
Wang et al. (uHCC) [28]	Retrospective/77	PD-1/PD-L1 inhibitor + targeted agents; OS (primary), PFS	MTV/TLG (ROC-derived cutoffs, MTV ≥ 39.65 cm3) outperform SUVmax for OS stratification	Volumetric metabolic burden better captures systemic disease aggressiveness and predicts OS more robustly than single-voxel SUVmax.	Single-center; heterogeneous combination regimens; ROC-derived thresholds risk overfitting; biology linkage to immune markers not established.
Kawamura et al. (HCC) [29]	Retrospective; screened 43, analyzed 20	Atezo/Bev; early PD within 12 weeks (e-PD)	TLR (tumor/liver) ≥2: e-PD 64% (7/11) vs. 11% (1/9); multivariable HR 7.153, P = 0.027	Baseline FDG TLR may flag high risk of very early failure → potential ‘early switch/escalation’ signal.	Strong selection bias; small N; TLR sensitive to liver background (cirrhosis/inflammation) and protocol.
Lee et al. (HCC) [30]	Retrospective/78	Atezo/Bev; PFS/OS	TLG cutoff 171 g: median PFS 3.3 months, OS 9.1 months in high-TLG; multivariable PFS HR 1.853 (P = 0.019), OS HR 2.686 (P = 0.003)	Metabolic tumor burden (TLG) provides reproducible-looking prognostic stratification under Atezo/Bev.	Retrospective; cutoff likely center-dependent; FDG uptake confounded by necrosis/inflammation and tumor differentiation.
Li et al. (early HCC, radiomics) [31]	Retrospective/80	Surgery; MVI (microvascular invasion) and DFS (disease free survival)	PET/CT radiomics model: MVI AUC 0.891; DFS nomogram C-index 0.831	Suggests PET radiomics may improve preop risk stratification beyond conventional metrics.	Classic radiomics risks: overfitting, segmentation/reconstruction dependence; external validation/standardization essential.

FDG, fluorodeoxyglucose; ICI, immune checkpoint inhibitor; LDH, lactate dehydrogenase; MTV, metabolic tumor volume; OS, overall survival; PD, progressive disease; PFS, progression-free survival; RECIST, Response Evaluation Criteria in Solid Tumors; ROC, receiver operating characteristic; TLG, total lesion glycolysis; TLR, tumor-to-liver ratio; uHCC, unresectable hepatocellular carcinoma.

In HCC specifically, FDG-PET has been explored in several clinically relevant contexts. In atezolizumab/bevacizumab-treated cohorts, baseline tumor-to-liver ratio (TLR) ≥2 was associated with a substantially higher rate of early progressive disease (within 12 weeks) and increased risk in multivariable modeling, while larger series further supported baseline TLG as an independent prognostic factor (e.g. TLG ≥ 171 g associated with shorter PFS/OS and remaining significant after adjustment). Beyond survival endpoints, a conversion-therapy study (PD-1 inhibitor plus lenvatinib) linked higher baseline metabolic measures (e.g. SUV-derived ratios) with poorer pathological response, albeit in a highly selected surgical subset, underscoring both the appeal of ‘hard’ pathological endpoints and the risk of selection bias. Together, these data suggest FDG-PET – particularly MTV/TLG and normalized ratios – functions primarily as a prognostic/early-failure risk marker complementary to immune-target and effector-function imaging, rather than a validated ICI-specific predictor [27–30].

PET/computed tomography radiomics and machine learning: imaging-derived surrogates of immune phenotypes and outcome risk

Radiomics/machine learning studies extend FDG-PET/CT from single-parameter readouts toward multifeature predictive signatures that aim to approximate immune biology or clinically relevant risk states. In non-small cell lung cancer, a retrospective multicohort study developed a PET/CT radiomics + clinical fusion model to predict tumor immune profiles, using a training cohort (n = 206) and an external public validation cohort (n = 39, The Cancer Imaging Archive), reporting good discrimination (AUC 0.86 in the test set and AUC 0.85 in external validation). Importantly, the target was an immune phenotype rather than ICI benefit, so its value for an HCC-focused review is primarily methodological transferability (i.e. imaging-based immune stratification is feasible) rather than direct clinical inference in HCC [31]. In early-stage HCC (Barcelona Clinic Liver Cancer 0–A), an FDG-PET/CT radiomics model (single retrospective cohort, n = 80) achieved an microvascular invasion-prediction AUC of 0.891 and constructed a DFS-related nomogram (C-index 0.831), supporting potential use for preoperative risk stratification rather than immunotherapy response prediction [31]. Across both settings, the dominant limitations are typical for radiomics: retrospective design, feature dimensionality relative to sample size, and strong sensitivity to acquisition/reconstruction/segmentation heterogeneity, meaning reported AUC/C-index values can inflate and may drift in external cohorts; rigorous multicenter standardization and prospective validation are prerequisites for clinical translation [31,32].

At the practice-framework level, an European Association of Nuclear Medicine symposium report/review synthesized how FDG-PET/CT can support immunotherapy response assessment and irAE recognition, emphasizing the limitations of RECIST (e.g. pseudo-progression/hyper-progression), summarizing response criteria (European Organisation for Research and Treatment of Cancer/PET Response Criteria in Solid Tumors and composite schemes such as PET/CT Criteria for Early Prediction of Response to Immune Checkpoint Inhibitor Therapy/PET Response Evaluation Criteria for Immunotherapy), and highlighting that the evidence base remains heterogeneous and often small-scale – useful for structuring discussion and acknowledging uncertainty rather than adopting any single universal cutoff [33].

⁶⁸Ga-prostate-specific membrane antigen PET in hepatocellular carcinoma: a qualitative response-assessment framework emphasizing interpretive consistency

Antony et al. addressed a prerequisite for any imaging biomarker to be clinically credible – reader consistency – by developing a four-tier qualitative ⁶⁸Ga-prostate-specific membrane antigen (PSMA) PET response framework for treatment assessment in HCC and benchmarking its interpretive agreement against cross-sectional criteria. Under an U.S. Food and Drug Administration Investigational New Drug screening context, 88 HCC patients underwent baseline PSMA PET, but only 13 met strict eligibility for paired baseline/posttreatment PSMA PET and CT/MRI response assessment (locoregional therapy 7 patients, 10 lesions; ICI 6 patients, 19 lesions). Three independent readers applied the qualitative scale and a dichotomized classification anchored to liver-background uptake, yielding near–‘almost perfect’ agreement (Gwet’s agreement coefficient: AC1 ≈ 0.90 for the dichotomized scale; AC2 ≈ 0.86 for the four-tier ordered scale), whereas mRECIST showed only moderate concordance; in the ICI subset, PSMA PET agreement remained superior to both mRECIST and RECIST 1.1 dichotomized assessments. Importantly, the authors framed this as a reliability study, establishing methodological readiness for subsequent outcome-anchored validation. The key limitations are the very small analyzable sample and strict selection (risk of selection bias), and the absence of direct validation against ORR/PFS/OS, meaning current evidence supports reproducible response categorization rather than proven clinical predictive validity; liver-background anchoring may also shift with liver function, inflammation, or acquisition differences and requires multicenter calibration [34].

Theranostics in nuclear medicine

Theranostics in nuclear medicine adopts a matched diagnostic–therapeutic paradigm in which a shared biological target is first quantified noninvasively and then exploited for radionuclide-based treatment. In HCC, this strategy is appealing because it enables patient-level and lesion-level selection based on in vivo target uptake, followed by targeted irradiation of uptake-positive disease, thereby integrating treatment planning, response monitoring, and adaptive management within a single workflow [35]. A commonly used approach is to pair diagnostic imaging with therapeutic radionuclides such as ¹⁷⁷Lu, which can deliver targeted β− irradiation and also support post-therapy imaging and dosimetry to confirm biodistribution and inform individualized dosing – features particularly relevant for advanced-stage disease where heterogeneous uptake is frequent [35].

In parallel, ⁶⁴Cu-based agents are increasingly used for PET-based target quantification and pharmacokinetic assessment, which can guide patient selection, optimize dosing, and facilitate rational pairing with therapeutic radionuclides in image-guided radionuclide therapy development [36]. Despite this promise, translation of theranostics in HCC remains limited by target heterogeneity, the need for standardized dosimetry and response criteria, and uneven access to radiochemistry and radionuclide therapy infrastructure. Accordingly, current evidence should be presented as evolving, with priority given to prospective studies that (i) link imaging uptake to clinically meaningful endpoints and (ii) establish reproducible selection thresholds across institutions.

Limitations, practical barriers, and future directions

Despite rapid progress, the current evidence base supporting nuclear medicine for predicting response in HCC remains methodologically heterogeneous and, in many domains, premature for routine clinical decision-making. First, a substantial proportion of published studies are retrospective, single-center, and underpowered, with small analyzable cohorts even when screening populations are larger. This design profile increases the risk of selection bias and limits the robustness of reported cutoffs (e.g. SUV- or TLR-based thresholds), which are often derived by within-cohort optimization rather than prospectively prespecified. Consequently, associations such as FDG metrics with PD-L1 expression or early progression should be interpreted as correlative and hypothesis-generating rather than definitive predictive biomarkers unless validated in prospective, outcome-anchored cohorts.

Second, biological heterogeneity in HCC complicates interpretation. Tumor differentiation, necrosis, background inflammation, and underlying liver disease can all influence FDG uptake, while immuno-PET signals may integrate contributions from tumor cells, immune infiltrates, and normal immune tissues. Therefore, imaging readouts should not be presented as direct surrogates for single-marker pathology (e.g. PD-L1 immunohistochemistry) without acknowledging spatial and cellular-source ambiguity and the potential for discordance between lesion-level biology and single-site biopsy results.

Third, inter-institutional variability is a major barrier to implementation. Differences in scanner platforms, reconstruction settings, uptake times, ROI/volume of interest definition, and partial-volume correction can materially shift quantitative metrics (SUV, MTV/TLG) and undermine cross-study comparability. This issue is particularly consequential for multicenter radiomics/artificial intelligence models, where acquisition and segmentation variability can lead to feature instability and performance drift. Standardization initiatives and reporting frameworks in nuclear medicine emphasize these pitfalls, yet prospective studies with harmonized protocols, quality control, and external validation remain limited.

Fourth, practical considerations may constrain adoption even when a signal is promising. Immuno-PET tracers labeled with long-lived radionuclides (e.g. ⁸⁹Zr) require complex radiochemistry, multiple imaging time points, and specialized infrastructure, which may limit accessibility and increase cost. Reimbursement uncertainty, patient burden from repeated visits, and radiation exposure trade-offs further influence feasibility. SPECT-based approaches can improve accessibility but may face lower spatial resolution and quantification challenges, again necessitating careful protocol standardization.

Finally, future studies should prioritize clinically actionable designs. Beyond demonstrating correlation, trials should test whether imaging-guided strategies improve decision-making – for example identifying early nonresponders for regimen escalation or switching, enriching patients for immunotherapy benefit, or differentiating pseudo-progression from true progression using predefined criteria. Where possible, imaging endpoints should be linked to ‘hard’ outcomes (ORR/PFS/OS) and mechanistic correlates (immune cell density/function, spatial transcriptomics, circulating biomarkers) to establish biological validity. Reliability work that quantifies interpretive consistency is also essential, because even a biologically informative biomarker is unlikely to translate if reader agreement is poor.

Conclusion

Nuclear medicine is emerging as an important component of precision management in HCC by enabling noninvasive, whole-body assessment of tumor metabolism, target availability, immune activation, and microenvironmental remodeling. While FDG-PET/CT and volumetric metabolic parameters may support baseline prognostic stratification and early failure-risk estimation, immuno-PET and functional immune imaging offer a conceptually stronger link to treatment mechanisms. Nevertheless, most current evidence is derived from heterogeneous, largely retrospective studies with limited sample sizes and protocol variability; thus, many findings should be viewed as promising but not yet practice-changing. The field’s near-term priority is rigorous prospective validation with harmonized imaging protocols and clinically meaningful endpoints, alongside pragmatic considerations of cost, accessibility, and workflow integration. With these steps, nuclear medicine-based biomarkers may evolve from exploratory correlates to validated tools that genuinely improve treatment selection and response-adaptive care in HCC.

Acknowledgements

This study is supported by Fujian Provincial Natural Science Foundation (No. 2023J011480) and the Institutional Research Project of Mengchao Hepatobiliary Hospital of Fujian Medical University (Grant No. 2025-LCYJ-05).

Conflicts of interest

There are no conflicts of interest.