Cutaneous melanoma (CM) is a common cancer in developed countries, with increasing incidence rates (1). Early-stage melanoma has a high 5-year survival rate (98.4%), but survival decreases at advanced stages, with rates of 63.6% for regional nodal involvement and 22.5% for metastatic disease (2). Advances in targeted therapies and immunotherapies, especially immune checkpoint inhibitors (ICI) targeting CTLA-4 (e.g., ipilimumab) (3) and PD-1 (e.g., nivolumab, pembrolizumab), have improved survival (4–6). The 5-year survival rate for CM patients treated with nivolumab is 39% (5), and the combination of nivolumab and ipilimumab shows a 5-year survival rate of 52% (6). However, the 12-month progression-free survival (PFS) rate with nivolumab+ipilimumab is 47.7%, 40 to 65% of metastatic melanoma do not respond to ICI therapies, and among responders, 43% develop secondary resistance after 3 years (7). Current challenges to improve the prediction of treatment response in melanoma include the investigation of the tumor microenvironment and tumor heterogeneity through “omics” approaches, including genomics and radiomics (8, 9).

Pre-therapeutic ¹⁸F-fluorodeoxyglucose positron emission tomography-computed tomography (FDG-PET/CT) is commonly used to stage advanced CM (10, 11). Studies have shown that tumor uptake, assessed through quantitative [e.g., standardized uptake value (SUV)] and volumetric [e.g., metabolic tumor volume (MTV), total lesion glycolysis (TLG)] parameters, is a prognostic factor for survival in various cancers, including CM (12). However, no definitive cut-off value has been established to select patients with poor prognosis based on these imaging parameters.

Radiomic approaches extract textural features (TFs) from imaging modalities such as PET to characterize tumor heterogeneity, which is significant in CM, with regions of hypoxia, necrosis, high cell proliferation, and angiogenic variability (13). While first-order features with SUV, MTV, and TLG capture basic tumor characteristics, second- and higher-order features model spatial voxel relationships, providing a more comprehensive analysis of heterogeneity (14). This method could help personalize treatment for CM, a cancer with poor prognosis, by enhancing therapy effectiveness. However, challenges in reproducibility and PET calibration remain, though advancements in harmonized protocols, artificial intelligence, and deep learning are improving these techniques (15).

Accurate early identification of patients at high risk of progression under ICI remains a major unmet clinical need in metastatic melanoma, as treatment intensification or de-escalation strategies are currently based on limited predictive tools (16, 17).

The aim of this study is to evaluate the prognostic value of a pre-treatment FDG-PET/CT-based radiomic model (MEL-RAD) for disease progression in a large cohort of patients with metastatic CM treated with ICI.

To our knowledge, this is the first study to validate a radiomic model based on FDG PET of ICI-treated CM patients in an external cohort.

The cohorts were marked by tumor heterogeneity, including cases with oligometastatic disease, but few other potential confounders, as all patients had an ECOG PS of 0 or 1 and no notable comorbidities. However, differences existed between the two cohorts: the testing cohort had a higher median age, a greater proportion of patients with ECOG PS 1 and a higher prevalence of elevated baseline LDH levels, suggesting a potentially more aggressive disease profile. Despite these differences, other clinical and biological characteristics were comparable between the cohorts.

MEL-RAD model was significantly associated with PFS in both cohorts (development and testing). However, for OS, the model showed a statistically significant association only in the development cohort.

The MEL-RAD model, developed by combining three radiomic features, demonstrated moderate discriminative ability in the development cohort, with an AUC of 0.74 (p < 0.0001). When using the 1y-PFS rate as the disease prevalence, the overall predictive accuracy of the model was estimated at 50.5%, indicating modest performance. An optimal probability threshold of 55% yielded a C-statistic of 0.63, with high Se (93.8%) but low Sp (31.9%). The corresponding PPV and NPV were 58.4% and 83.4%, respectively. External validation of the MEL-RAD model in an independent cohort (testing) confirmed these findings, with a C-statistic of 0.62, Se of 95.7%, Sp of 27.8%, PPV of 45.8%, and NPV of 90.9%. Although the low specificity indicates a substantial rate of false positives, the high Se and elevated NPV suggest that the model is effective in identifying patients who are unlikely to experience disease progression.

Predicting prognosis prior to treatment initiation remains essential for optimizing the management of patients with CM (8, 9, 26). Only three studies have investigated the predictive value of TFs extracted from FDG-PET/CT, but these were based on small sample sizes. Most of the other studies focused primarily on first-order PET indices (27–29). Baseline SUVpeak, MTV, and TLG have also been reported as promising predictors of response to ICI treatment in CM (30), but no clear thresholds have been established to stratify patients into prognostic subgroups. Emerging evidence suggests that MTV may predict survival, although its robustness varies, and further large studies are needed to confirm and validate MTV thresholds for clinical use. Notably, approximately 55% of CM cases show intrinsic resistance to anti-PD-1 inhibitors (with around 40% showing innate resistance to CTLA4 + PD1 inhibitor combination), with 25% of initial responders developing resistance within 2 years (31). In our study, these first-order TFs were excluded from the predictive model due to their limited prognostic value, as they measure global tumor characteristics without analyzing complex relationships between voxels or the internal tumor structures. In contrast, Higher-order TFs, which were found to be significant in our model, provide more complex information reflecting tumor heterogeneity and indicating rapid tumor proliferation (14). The three selected higher-order radiomic features likely capture different aspects of intra-tumoral heterogeneity, which may explain their predictive value. LHH_GLRLM_GrayLevelVariance reflects variability in gray-level run lengths, HLL_GLCM_ClusterProminence emphasizes asymmetry and peakedness in spatial gray-level patterns, and HHH_GLDM_GrayLevelVariance quantifies local intensity dependencies (14, 32–34). Together, these features provide complementary information on tumor texture complexity, potentially underlying the MEL-RAD model’s ability to stratify patient risk. Indeed, despite the limited data available in the literature for these features in PET imaging, GrayLevelVariance measures the dispersion of gray intensities in a given image area. A high variance indicates greater intensity changes, reflecting significant tumor heterogeneity and potential aggressiveness. In Ferrer-Lores’ study on diffuse large B-cell lymphoma, this feature was predictive of a therapeutic response to chemotherapy when its value was low (35). Similarly, Cluster Prominence (CP) is a measure of the asymmetry of a given distribution. High CP values indicate low image symmetry. In medical imaging, low CP values represent a smaller peak for the image gray-level value, and usually, the gray-level difference between the forms is small. This reflects heterogeneity, as the lower the value, the greater the heterogeneity and potential aggressiveness. In the study by Huang et al., this feature was predictive of breast cancer recurrence at 1 year following treatment with chemotherapy, surgery, or radiotherapy (36).

Our model may therefore influence the treatment strategy of metastatic CM under ICI, as it showed superior risk stratification compared to established first-order TFs (9). We proved a risk reduction of almost 20% of DP. Moreover, patients with a MEL-RAD probability >55% could benefit from intensified treatment (anti-PD1 + anti-CTLA4 or anti-LAG3 for PD-L1 <1%). This is consistent with evidence showing improved outcomes and reduced hyperprogression with combination therapy (24). Conversely, those with a MEL-RAD probability ≤55% may be better suited for anti-PD1 monotherapy, which offers fewer adverse events (AEs) (37).

Our clinicobiological results were not statistically significant, likely due to limited statistical power stemming from the relatively small sample size. However, there was a trend toward significance for brain metastases in relation to PFS, which aligns with findings in the literature. In the study by Moyers et al. (38), the presence of brain metastases was associated with shorter PFS compared to their absence. Regarding OS, no significant factors were identified, although elevated baseline LDH levels showed a trend toward significance, consistent with previously published data (39).

We used a robust and rigorous methodology in this radiomic study, which is essential to ensure reproducibility and clinical applicability (14). One of the major challenges in radiomics is therefore to ensure reproducible tumor delineation across studies and different centers and clinical settings, as contour variations can affect the analysis. In our study, we used a validated segmentation technique based on the PERCIST approach for treatment response assessment in delineating the target tumor with a threshold of 1.5 x mean liver SUV in a 3-cm spherical VOI in the right lobe + 2 SD (20). However, this method shares the limitations of absolute SUV-based thresholds, particularly in accounting for intratumoral heterogeneity (40). While the gradient-based method excludes necrotic areas and offers an excellent reproducibility (41), it is not yet widely available with all commercially software (42). To date, there is no consensus on the methodology for defining the metabolic threshold for FDG-PET/CT segmentation (43). Some studies used a fixed SUV threshold for tumor delineation (27, 28, 44), while our study normalized the threshold for each patient using liver uptake as a reference (20). Our semi-automated method combined with a manual correction enables for rapid tumor delineation followed by refinement for greater accuracy and consistency. This hybrid approach improves reproducibility while accommodating complex tumor shapes and heterogeneous regions. However, we acknowledge that our segmentation pipeline does not provide the evaluation of inter-reader variability. Another important challenge in radiomics is the extraction of reliable features from small-volume lesions, which are prone to high noise levels and low signal-to-noise ratios. Extracting radiomic parameters from these small lesions carries the risk of obtaining inaccurate or unreliable features, potentially leading to biased results (45). However, we have demonstrated the feasibility of this approach by using an established feature extraction methodology that has already been externally validated through rigorous testing (46). Our study provides strong evidence that even when integrating small-volume lesions, it is possible to extract meaningful radiomic features that are both robust and reproducible, provided the appropriate methodology is applied. Finally, our strength was to build our prognostic model MEL-RAD from a development cohort and test it in an external validation cohort, which is recognized to improve its applicability in different clinical settings (47). Moreover, the use of the neuroCombat harmonization method (22) in our study minimized the heterogeneity of the multicenter data without compromising the inter-patient variability (48).

Our study has several limitations. First, the cohort is heterogeneous, including several patients with pauci-metastatic disease or with BRAF-mutation or on anti-PD1 monotherapy alone. However, this reflects real-life and current practice. Second, the retrospective nature of the study introduces a risk of selection bias. Third, the follow-up period was approximately 16 months, and although this is sufficiently powered for significant PFS results, a longer-term evaluation would be valuable particularly for OS analysis. Fourth, certain characteristics were not assessed such as the Fitzpatrick skin phototype. Fifth, The MEL-RAD model’s performance profile indicates that, while it may effectively exclude patients at low risk of progression (functioning as a “rule-out” model), its capacity to accurately identify those at high risk remains limited. Finally, the study was conducted in two centers, which is a methodological challenge (49). As mentioned above, we used harmonization procedures between the different machines in order to limit the variability due to the different spatial resolution of the PET systems and image reconstruction methods (50). This ensured the robustness of the model for both internal and external validation.

In conclusion, despite the inherent limitations commonly encountered in predictive modeling (51), our results were highly significant and demonstrated the promising predictive potential of FDG-PET/CT radiomics in metastatic CM patients treated with immunotherapy. This model offers encouraging prospects for therapeutic personalization as a biomarker (52) and lays the groundwork for future multimodal approaches that integrate radiomic features with clinical data in larger cohorts. Such combined strategies hold the potential to ultimately enhance prognostic stratification and guide treatment decisions more effectively in this patient population.