PET Radiomics Signatures and Artificial Intelligence for Decoding Immunotherapy Response in Advanced Cutaneous Squamous Cell Carcinoma: A Retrospective Single-Center Study

Featured Application

Radiomic analysis of baseline [¹⁸F]FDG PET/CT scans may offer a non-invasive tool to predict immunotherapy response and tumor grade in patients with advanced cutaneous squamous cell carcinoma. This approach could support clinical decision making by identifying likely responders prior to treatment initiation and tailoring management strategies based on tumor differentiation.

Abstract

The aim of this study was to develop a baseline [¹⁸F]FDG PET/CT model to predict immunotherapy response in advanced cutaneous squamous cell carcinoma (cSCC) and noninvasively determine tumor grade, thereby enhancing early patient stratification. We retrospectively analyzed 59 patients with histologically confirmed advanced cSCC submitted to immunotherapy with cemiplimab. All underwent [¹⁸F]FDG PET/CT at baseline and after approximately 12 weeks. Clinical response was assessed through PET findings integrated with clinical and dermatological evaluation, and patients were classified as responders (complete/partial metabolic response or stable disease) or non-responders (progression or toxicity-related discontinuation). Tumors were also classified as low to intermediate (G1–G2) or poorly differentiated (G3). Machine learning models (Random Forest and Extreme Gradient Boosting) were trained to predict treatment response and tumor grade. Clinical benefit was observed in 46/59 patients (77.9%), while 13 (22.1%) were non-responders. Histology showed 64.4% (n = 38) G1–G2 and 35.6% (n = 21) G3 tumors. The PET-based model best predicted clinical benefit (AUC = 0.96, accuracy = 91% cross-validation; AUC = 0.88, accuracy = 82% internal validation). For tumor grade prediction, the CT-based model achieved a higher AUC of 0.80 (accuracy 73%), whereas the PET-based model reached an AUC of 0.78 but demonstrated a slightly higher accuracy of 77%. Radiomic analysis of baseline [¹⁸F]FDG PET enables the discriminative prediction of immunotherapy response and tumor grade in advanced cSCC, with PET-based models outperforming CT-based ones.

1. Introduction

Cutaneous squamous cell carcinoma (cSCC) is the second most common form of non-melanoma skin cancer, accounting for approximately 20% of all skin malignancies [1]. While most cases are diagnosed early and managed effectively with surgery or radiotherapy, about 2–5% progress to locally advanced or metastatic disease, which carry significant morbidity and a poor prognosis (median overall survival ≤ 20 months) [2]. These advanced forms of cSCC are associated with significant morbidity, limited therapeutic options, and a poor prognosis, with median overall survival rarely exceeding 15–20 months in the absence of effective systemic treatment [3].

Cancer immunotherapies—including immune checkpoint inhibitors, CAR-T cells, oncolytic viruses, and vaccines—work by reactivating antitumor immune surveillance, enhancing T-cell activation, or reshaping the tumor microenvironment to eliminate malignant cells [4,5,6]. However, these approaches face critical limitations: low overall response rates and primary or acquired resistance driven by an immunosuppressive tumor milieu; risks of immune-related adverse events (e.g., cytokine release syndrome); and challenges in drug delivery, manufacturing complexity, and high costs [6]. Ongoing research aims to overcome these hurdles via combination regimens, next-generation engineering (e.g., Fc-optimized antibodies and bispecifics), and novel small-molecule and cell-based strategies to improve efficacy and safety.

Immunotherapy with cemiplimab, a fully human monoclonal antibody targeting programmed cell death protein 1 (PD-1), has shown 40–50% response rates in advanced cSCC [7,8,9], but nearly half of patients either fail to respond or develop resistance.

Currently, few validated tools are available to anticipate the response to immunotherapy in cSCC [10,11]. Conventional metrics, such as the maximum standardized uptake value (SUVmax), metabolic tumor volume (MTV), and total lesion glycolysis (TLG), calculated with positron emission computed tomography (PET/CT), provide only a macroscopic tumor burden estimate, missing spatial heterogeneity [12].

Emerging evidence suggests that radiomics, the high-throughput extraction and analysis of quantitative imaging features from medical images, may offer a deeper characterization of tumor phenotypes [13,14]. Radiomic features capture subtle variations in voxel intensity, texture, shape, and spatial distribution that are often imperceptible to the human eye. Applied to [¹⁸F]FDG PET/CT imaging, radiomics could enable the identification of imaging biomarkers predictive of immunotherapy response by capturing the biological heterogeneity associated with tumor aggressiveness and immune evasion [15].

In this study, leveraging our prior handcrafted PET/CT radiomics experience [16,17] —a supervised, controlled feature-engineering approach—we aim to develop a baseline [¹⁸F]FDG PET/CT model to predict immunotherapy response in advanced cSCC and noninvasively determine tumor grade, thereby enhancing early patient stratification.

2. Materials and Methods

2.1. Study Design and Patient Population

This retrospective observational study was approved by the local institutional ethics committee (protocol no. 0168351/2020). All consecutive subjects with histologically proven cSCC, submitted to PD-1 blocker cemiplimab and subjected to PET/CT scans between January 2021 and August 2023, were considered potentially eligible for enrollment. Eligible patients were identified from our institutional database and met the following inclusion criteria: (1) histologically confirmed cSCC; (2) age ≥ 18 years; (3) treatment with cemiplimab (350 mg intravenous infusion every three weeks); (4) availability of baseline [¹⁸F]FDG PET/CT scan (PET-1) performed within 4 weeks prior to therapy initiation, and at least one follow-up scan (PET-2) performed approximately 12 weeks after the first dose (acceptable interval: 11–14 weeks); (5) presence of at least one FDG-avid lesion on PET-1; (6) complete clinical and dermatological follow-up for at least 6 months after PET-1. Exclusion criteria included the absence of FDG-avid lesions on baseline PET or incomplete clinical/imaging data. This study was performed in accordance with the ethical standards in the 1964 Declaration of Helsinki and its later amendments.

2.2. PET/CT Imaging

All patients underwent [¹⁸F]FDG PET/CT scans acquired with the same digital Biograph Vision 450 PET/CT scanner (Siemens Healthineers, Erlangen, Germany) following standard clinical protocols described elsewhere [12]. Patients fasted for at least 6 h prior to the intravenous administration of 3.7 MBq/kg of [¹⁸F]FDG. Blood glucose levels were confirmed to be <200 mg/dL prior to tracer injection.

PET/CT acquisition began 60 min post-injection, covering the region from the skull base to the proximal thighs (or extended when necessary). CT parameters included a slice thickness of 1.0 mm, a pitch factor of 1, and maximum 120 keV and 90 mAs with CARE Dose and CARE kV modulation. PET images were acquired in 3D mode using continuous bed motion (FlowMotion™, York, UK) at a speed of 0.9 mm/s and reconstructed using the TrueX + TOF algorithm (2 mm Gaussian filter, matrix size 440 × 440). Attenuation correction was applied using a low-dose CT scan.

PET/CT image interpretation was independently conducted by two board-certified nuclear medicine specialists (L.F. and R.P.). Image review was carried out using the Advantage Workstation 4.7 (GE HealthCare, Milwaukee, KY, USA). Any focus of radiotracer uptake exceeding the background and not attributable to physiological distribution was considered suspicious for malignancy. For each patient, both the anatomical locations and the number of abnormal FDG-avid lesions were systematically recorded. The primary cutaneous tumor and, when present, metastatic localizations were identified directly on PET images.

2.3. Imaging Post-Processing and Feature Extraction

The following parameters were calculated on the baseline PET/CT (PET-1): the number of FDG-avid lesions, maximum and mean standardized uptake value (SUVmax and SUVmean), total MTV, and TLG. In each patient, the primary cutaneous tumor was segmented by applying a threshold of 40% of the SUVmax within the lesion’s bounding box using dedicated software (PET VCAR; Advantage Workstation 4.7, GE Healthcare), in accordance with standard practice (Figure S1). In cases of multiple skin lesions, the one with the highest uptake value (SUVmax) was classified as the index lesion and segmented.

Radiomics features (RFts) were extracted from the VOIs of tumors using the Pyradiomics package (version 3.1.0). Supplementary Text S1 shows the parameters used for radiomics feature extraction from CT and PET images. In our pipeline, to control multicollinearity, among radiomic features, we applied a two-step filter-based selection strategy. The features significantly associated with outcome (best clinical response or tumor grade) were identified using the Mann–Whitney U-test with p-value thresholds of 5.8 × 10⁻⁵ and 0.05 for primary and secondary aim points, respectively. Subsequently, to reduce redundancy, pairwise Spearman rank correlation coefficients were calculated among the selected features. These were the statistical criteria for labeling a feature as robust. Inter-observer variability was minimized through the use of a standardized semi-automatic segmentation approach. Specifically, each primary tumor was segmented using a threshold of 40% of SUVmax within the lesion’s bounding box, as implemented in PET VCAR (Advantage Workstation 4.7, GE Healthcare). This threshold-based method, normally used in clinical practice, limits variability due to operator-dependent contouring.

Complete details of the selected features and their associated p-values are provided in the Supplementary Material (Tables S1–S4).

Features were extracted from the original and wavelet-filtered images, resulting in a set of quantitative biomarkers for each 3D VOI in both the PET and CT datasets. Radiomic features can be classified into three classes: shape-based features, statistical features (including histogram-based and texture-based features), and transform-based features. The Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Run-Length Matrix (GLRLM), Gray-Level Size-Zone Matrix (GLSZM) and Gray-Level Distance-Zone Matrix (GLDZM), and, finally, Neighborhood Gray-Tone Difference Matrix (NGTDM) and Neighborhood Gray-Level Dependence Matrix (NGLDM) were calculated.

For the wavelet-filtered images using Coiflet1, only second-order texture features were extracted.

2.4. Assessment of Immunotherapy Response

Therapeutic response was evaluated approximately 12 weeks after treatment initiation (PET-2) using PET Response Criteria in Solid Tumors (PERCIST). Lesions were categorized as showing complete metabolic response (CMR), partial metabolic response (PMR), stable metabolic disease (SMD), or progressive metabolic disease (PMD) [9].

The best clinical response (BCR) to immunotherapy was determined within 6 months from treatment initiation and classified as clinical benefit (CB: including CMR, PMR, or long-term stable disease) or no clinical benefit (NCB: confirmed progression or treatment discontinuation due to toxicity or clinical worsening). The BCR was assessed through the integration of clinical examination (including dermatologic and dermatoscopic evaluation), laboratory tests, and imaging (PET/CT, ultrasound, and/or MRI). Histological confirmation of progression was obtained where appropriate.

2.5. Follow-Up

Patients were monitored with clinical evaluations and laboratory tests on a monthly basis. Progression-free survival (PFS) was defined as the time between the PET scan and the occurrence of locoregional or distant relapse. Each case was discussed during the weekly multidisciplinary tumor board meeting held at our hospital. Imaging studies (ultrasonography, CT, PET/CT, or MRI) were performed, in addition to the clinical and laboratory follow-up, whenever deemed appropriate by the multidisciplinary team, which included dermatologists, oncologists, nuclear medicine physicians, and radiologists.

2.6. Model Building

Machine learning (ML) predictive models were developed using Orange Data Mining and Python package (version 3.8). Since our study was conducted using a single PET/CT scanner and a uniform acquisition and reconstruction protocol across all patients, no feature harmonization techniques were needed.

To address class imbalance in both the tumor grade and best clinical response (BCR) datasets, SMOTE (Synthetic Minority Oversampling Technique) was applied. Separate ML models were built for each target (tumor grade and BCR) and for each imaging modality (PET and CT) using robustly selected features. Two classifiers were employed: Extreme Gradient Boosting (XGBoost) and Random Forest (RF). Each dataset—PET and CT for both BCR and tumor grade—was independently split into a training cohort (70% of the data) and an internal validation cohort (30%). A 10-Fold cross-validation (10-Fold CV) was performed on the training sets, and performance metrics were averaged across the folds. The remaining 30% of each dataset was used for internal validation. Model performance was evaluated using several metrics, including the Area Under the Receiver Operating Characteristic Curve (AUC), classification accuracy (CA), precision (PRE), sensitivity (SEN), specificity (SPE), and true-positive (TP) and true-negative (TN) counts. The Receiver Operating Characteristic (ROC) analysis was performed, and the subsequent ROC plots were built. In total, four different ML models were trained and validated, namely, BCR_CT, BCR_PET, GRADE_CT, and GRADE_PET.

A schematic flowchart that illustrates the full machine learning pipeline, including data preprocessing, feature selection, model training, and evaluation, is depicted in the Supplementary Material (Figure S2).

2.7. Statistical Analysis

Descriptive statistics were used to summarize clinical and imaging data. The association between radiomic features and outcome variables (tumor grade and BCR) was evaluated using a univariate nonparametric statistical test. The primary endpoint of this study, i.e., evaluating whether PET- and CT-based radiomic features can predict clinical response to cemiplimab (BCR), was investigated using the Mann–Whitney U-type test implemented with Bonferroni correction. Features with corrected p-values below this adjusted Bonferroni threshold (5.8 × 10⁻⁵) are considered statistically significant.

The secondary endpoint, i.e., the investigation into whether radiomic features can noninvasively predict histopathological tumor differentiation, distinguishing low-to-intermediate-grade lesions from poorly differentiated cSCC, was conducted using a double filter features selection algorithm. The algorithm selected robust features using the Mann–Whitney U test and Spearman rank correlation (p < 0.05).

3. Results

3.1. Patient Population

Out of an initially recruited 94 patients, 59 were ultimately enrolled in the study, while 35 were excluded due to either the lack of a follow-up PET-CT scan or incomplete clinical history. Clinical and demographic characteristics of the enrolled patients are summarized in

Table 1. Patients’ clinical and demographic characteristics.

Characteristics	N (Percentage)
Patients	59
Male sex	51 (86.4%)
Median age y (interquartile range)	86 (84–89)
ECOG
0–1	47
≥2	12
Tumor Location
Head/neck	52 (88.1%)
Trunk	4 (6.7%)
Lower limbs	3 (5%)
Tumor grade
G1–G2	38 (64.4%)
G3	21 (35.6%)
Metastatic patients	8 (13.5%)
Nodal metastases	6 (10.1%)
Distant metastases	2 (3.3%)
Immunotherapy with respect to other therapies
As 1st line	46 (77.9%)
After upfront surgery	12 (20.3%)
After upfront RT	1 (1.7%)
PET-based response after 12 weeks (PERCIST)
CMR	5 (8.4%)
PMR	39 (66.1%)
SMD	2 (3.3%)
PMD	13 (22%)
Best clinical response (BCR) to immunotherapy
CB	46 (77.9%)
NCB	13 (22.1%)
Immunotherapy-related toxicity
Cardiovascular toxicity	1 (1.7%)
Therapeutic-switch after immunotherapy failure
Chemotherapy	2 (3.3%)
Radiotherapy	3 (5%)
Palliative treatments	8 (13.5%)
Progression-free survival (median, months)	14.7

CMR: complete metabolic response; PMR: partial metabolic response; SMD: stable metabolic response; PMD: progressive metabolic disease; CB: clinical benefit; NCB: no clinical benefit.

.

The final cohort was predominantly male (86.4%), with a median age of 86 years. Regarding performance status, 47 patients were classified as ECOG 0–1 and 12 as ECOG ≥ 2, indicating overall good functional capacity across the study population. The tumor was primarily located in the head and neck region (88.1%), with fewer cases arising in the trunk (6.7%) and lower limbs (5%). Histopathological analysis demonstrated that 64.4% of patients had low-grade tumors (G1–G2), while 35.6% exhibited high-grade (G3) disease. Metastatic involvement was observed in a minority of cases, with nodal metastases in 10% and distant metastases in 3.3% of patients. In terms of treatment, immunotherapy was used as a first-line approach in 77.9% of patients; additionally, 20.3% had undergone previous surgical intervention and 1.7% had received prior radiotherapy. Given the relatively homogeneous distribution of age (median 86 years, interquartile range: 84–89), ECOG (predominantly 0–1), and low metastatic burden in our cohort, we deliberately focused on baseline imaging features to isolate the independent predictive value of radiomics without introducing potential confounders.

3.2. Response to Immunotherapy

A clinical benefit (CB) from immunotherapy was observed in 77.9% of cases, whereas 22.1% did not experience such a benefit (NCB). At PET-2, 46 patients were classified as responders (CMR, PMR, or SMD), while 13 were deemed non-responders due to evidence of progressive metabolic disease. Figure 1 shows an example of a patient with SCC who experienced a CB after immunotherapy and also showed a metabolic response on PET-2. In contrast, Figure 2 illustrates the condition of a patient with metabolic progression on PET-2, classified as having NCB.

Among those classified as responders at PET-2, all also demonstrated a clinical benefit based on clinical and dermatological evaluations. For the 13 non-responders, the absence of clinical benefit was established by dermatological evaluation along with histological confirmation in seven cases and cytological confirmation in one case (eight cases in total), with the remaining five cases being determined solely by objective assessment. Among patients who failed to respond to immunotherapy, subsequent therapeutic strategies included chemotherapy (n = 2), radiotherapy (n = 3), and palliative treatments (n = 8). The median progression-free survival was 14.7 months.

3.3. Radiomic Analysis

A total of 112 original texture features and 744 wavelet-based texture features were extracted, resulting in 856 RFts per case. To address class imbalance, the SMOTE oversampling technique was applied, resulting in two balanced datasets: the BCR dataset with 92 samples (46 per class) and the tumor grade dataset with 76 samples (38 per class), respectively. The results are presented in two sections: BCR feature selection and tumor grade feature selection.

In the BCR analysis, the feature selection process identified 7 optimal RFts for the BCR_CT model and 86 for the BCR_PET model. All seven features selected for the CT-based model were derived from high-frequency wavelet transformations. In contrast, among the features selected for the PET-based model, 85 were wavelet-based (spanning both low- and high-frequency components), while 1 belonged to the second-order texture category: GLSZM Large Area High Gray Level Emphasis.

Regarding tumor grade prediction, statistical testing led to the selection of 34 and 24 robust RFts for the CT and PET models, respectively. Among the CT-derived features, two were original (Shape Maximum 2D Diameter Row and First Order Minimum), while the remaining 32 were wavelet-based. For the PET-based model, two of the 24 selected features—Shape Maximum 2D Diameter Row and GLDM Gray Level Non-Uniformity—were non-wavelet, with the rest derived from wavelet transformation.

Complete details of the selected features and their associated p-values are provided in the Supplementary Materials.

3.4. Performance of the Machine Learning Models

Extreme Gradient Boosting and Random Forest Classifiers were trained using 70% of the datasets resulting from the features selection steps. The two classifiers were evaluated in terms of several metrics, and the results on the training dynamics are shown using the 10-Fold CV; the remaining 30% of the dataset was considered for internal validation purposes.

Considering the ML models developed to predict the best clinical response (BCR_CT and BCR_PET), the results from 10-Fold CV were encouraging, with all AUC values exceeding 0.89. The best-performing model in this phase was the XGBoost classifier trained on BCR_PET-derived RFts, which achieved an AUC of 0.97 and CA = 0.91 (

Table 2. Performances scores in 10-Fold CV for each BCR ML Model.

	Training 10FoldCV	AUC	CA	PRE	SEN	SPE	TP	TN
BCR_CT Model	RF	0.91	0.87	0.87	0.87	0.87	82.90%	90.20%
	XGB	0.89	0.88	0.88	0.88	0.88	85.00%	90.40%
BCR_PET Model	RF	0.96	0.91	0.91	0.91	0.92	93.60%	88.90%
	XGB	0.97	0.91	0.91	0.91	0.91	91.80%	90.70%

AUC = Area Under the Curve; BCR = best clinical response; CA = classification accuracy; PRE = precision; RF = Random Forest; SEN = sensitivity; SPE = specificity; XGB = Extreme Gradient Boosting; TP = true positive; TN = true negative.

). In internal validation, all models continued to show satisfactory performance, with AUC values above 0.71. Notably, the RF BCR_PET model achieved an AUC of 0.88, a CA of 0.82, a TP rate of 68.80%, and a TN rate of 94.10%. The models demonstrated particularly strong discriminative power for the negative class (i.e., non-responders), as reflected by the high TN rate. The ROC curves obtained from the 10-Fold CV and internal validation are shown in Figure 3.

In the tumor-grade classification task, the CT-based models achieved AUCs of 0.78 with RF and 0.83 with XGBoost during 10-Fold CV. Both models reached a CA of 76% and exhibited comparable performance in terms of precision (0.77 vs. 0.76), sensitivity (0.76 for both), and specificity (0.74 vs. 0.75). The PET-based models yielded AUCs of 0.77 with RF and 0.79 with XGBoost, with XGBoost outperforming Random Forest in accuracy (78% vs. 70%) and demonstrating slight improvements in precision, sensitivity, and specificity (

Table 3. Performances scores in 10-Fold CV for each GRADE ML model.

	Training 10-Fold CV	AUC	CA	PRE	SEN	SPE	TP	TN
GRADE_CT Model	RF	0.78	0.76	0.77	0.76	0.74	80.00%	73.50%
	XGB	0.83	0.76	0.76	0.76	0.75	77.30%	75.00%
GRADE_PET Model	RF	0.77	0.70	0.71	0.70	0.69	71.40%	69.70%
	XGB	0.79	0.78	0.78	0.78	0.77	78.30%	77.40%

AUC = Area Under the Curve; CA = classification accuracy; PRE = precision; RF = Random Forest; SEN = sensitivity; SPE = specificity; XGB = Extreme Gradient Boosting; TP = true positive; TN = true negative.

). In the internal validation cohort, XGBoost maintained its superiority in the CT setting, achieving an AUC of 0.80 compared to 0.70 for Random Forest, along with higher classification accuracy (73% vs. 55%), precision (0.89 vs. 0.67), and specificity (0.89 vs. 0.67). In the PET setting, both models demonstrated robust performance, with AUCs of 0.75 for RF and 0.78 for XGBoost. XGBoost offered modest gain in precision (0.81 vs. 0.78) and specificity (0.81 vs. 0.77), as shown in

Table 4. Performances scores in the validation step for each ML model.

	Internal Validation	AUC	CA	PRE	SEN	SPE	TP	TN
BCR_CT Model	RF	0.75	0.67	0.79	0.58	0.79	78.60%	57.90%
	XGB	0.71	0.73	0.78	0.74	0.71	77.80%	66.70%
BCR_PET Model	RF	0.88	0.82	0.85	0.82	0.86	68.80%	94.10%
	XGB	0.87	0.76	0.82	0.76	0.83	61.10%	93.30%
GRADE_CT Model	RF	0.70	0.55	0.67	0.46	0.67	66.70%	46.20%
	XGB	0.80	0.73	0.89	0.62	0.89	88.90%	61.50%
GRADE_PET Model	RF	0.75	0.77	0.78	0.77	0.77	83.30%	70.00%
	XGB	0.78	0.77	0.81	0.77	0.81	90.00%	66.70%

AUC = Area Under the Curve; BCR = best clinical response; CA = classification accuracy; PRE = precision; RF = random forest; SEN = sensitivity; SPE = specificity; XGB = Extreme Gradient Boosting; TP = true positive; TN = true negative.

. The ROC curves obtained from the 10-Fold CV and internal validation are shown in Figure 4.

The scores for each validated ML model are graphically represented in the radial plots shown in Figure 5. In addition, Figure 6 graphically compares the two best-performing models in the BCR prediction task, with RF learner for CT data and XGBoost learner for PET data. For tumor grade prediction, XGBoost emerged as the best-performing learner, with the PET-based model outperforming the CT-based model.

Performance metrics from the 10-Fold CV for each BCR and GRADE ML model are summarized in Table 2 and Table 3, respectively, and visualized in Figure S3. Table 4 reports the internal validation results for all models.

4. Discussion

In this exploratory cohort of 59 patients, the PET-based radiomic models achieved cross-validated AUCs up to 0.97 for immunotherapy response and 0.85 for tumor grade, with internal validation AUCs of 0.88 and 0.80, respectively. These results illustrate the potential of handcrafted radiomics to outperform conventional PET metrics in predicting both treatment benefit and histopathologic differentiation [18,19].

Advanced cSCC is a rare, challenging disease of predominantly elderly, frail patients with limited systemic options. While cemiplimab yields 40–50% response rates [8,9], there remains an urgent need for noninvasive, image-based biomarkers to guide therapy selection and avoid unnecessary toxicity in non-responders. By selecting a homogeneous elderly cohort, we aimed to assess the intrinsic value of baseline radiomic heterogeneity; adding clinical covariates in larger, more diverse populations will be an important future step to mirror real-world decision making.

Our PET-based radiomic models achieved cross-validated AUCs up to 0.97 and internal validation AUCs up to 0.88, clearly outperforming CT-based models. This highlights how metrics of metabolic heterogeneity capture immune-relevant tumor biology beyond conventional volumetric measures such as SUVmax, MTV, and TLG. A key strength of our cohort is that 77.9% of patients received immunotherapy as first-line treatment, reflecting real-world practice in which cemiplimab is increasingly adopted upfront for advanced cSCC. In this context, imaging and laboratory biomarkers to stratify patients by likelihood of benefit are urgently needed.

We also assessed a secondary endpoint: the ability of PET radiomics to infer tumor grade “in vivo.” Accurate noninvasive grading is invaluable because histological grade reflects underlying biology—proliferation, differentiation, stromal interactions—and correlates with aggressiveness and evolutionary potential [20,21]. Whereas conventional biopsy may undersample heterogeneous high-grade regions, radiomic texture features survey the entire lesion (and multifocal sites), capturing spatial heterogeneity and temporal evolution [22]. Our grade-prediction models yielded AUCs of 0.85 (cross-validated) and 0.80 (internal validation), indicating that baseline metabolic heterogeneity encodes histopathologic grade. Such noninvasive predictions could inform treatment intensity and surveillance, and—when combined with immunotherapy-response models—generate a multiparametric signature of tumor biology that guides both prognosis and therapeutic sensitivity [23]. However, in this regard, while an AUC ≥ 0.8 is generally considered the minimum for a clinically useful screening biomarker, confirmatory testing and prospective calibration will be required before routine use of tumor-grade radiomics.

Regarding laboratory biomarkers, elevated or rising IL-6 levels have been linked to poorer outcomes in cemiplimab-treated cSCC [24]. A recent translational study identified early immune and transcriptomic changes as predictive markers: responders showed increased intratumoral B and CD8+ T cells, plus early downregulation of IL1β and IL8 (also mirrored in peripheral blood). Moreover, PD1+ regulatory T cells declined in responders but rebounded in non-responders, highlighting IL8 and PD1+ Tregs as promising early indicators of response [25].

Radiomics complements laboratory data by providing a noninvasive, image-based means to enrich for likely responders—especially relevant in our predominantly elderly cohort, for whom toxicity and quality-of-life concerns are paramount [26]. By identifying complex texture patterns on baseline PET scans, our models could support clinical decision making at therapy initiation, potentially contributing to a composite predictive score alongside clinical and laboratory variables.

Although PET/CT is not yet standard for assessing immunotherapy response in cSCC, McLean et al. [27] reported high discordance between metabolic complete response (CMR) on FDG-PET and anatomical response per RECIST1.1 on CT/MRI in cemiplimab-treated patients. In their retrospective Australian cohort (n = 15; median age 73 years, 93% male), 73% achieved CMR on PET, yet only one of these met the RECIST1.1 criteria for complete response, suggesting that PET may detect deeper, immune-mediated tumor eradication not captured by size-based measures.

Our work moves beyond simple CMR/PMR categories by applying high-throughput quantification of handcrafted radiomic features. This approach uncovers subtle spatial and textural patterns—dimensions that may more accurately predict long-term benefit than binary response metrics.

In our prior study of 25 cSCC patients treated with cemiplimab, baseline MTV and TLG both correlated significantly with event-free survival (p < 0.05), and a lack of metabolic response at 12 weeks was linked to markedly shorter survival (7.2 ± 1 vs. 20.3 ± 2.3 months). On multivariate analysis, metabolic response at 12 weeks remained the sole independent predictor. While these conventional metrics are prognostically valuable, they miss intratumoral heterogeneity and spatial complexity. By integrating radiomic features with MTV and TLG, we aim to build a more comprehensive, multiparametric signature of tumor aggressiveness and immunogenicity [12].

In our cohort, PET-based radiomics outperformed CT-based radiomics for the prediction of immunotherapy response. This may stem from greater feature variability—and thus more critical robustness assessment—in CT radiomics when lesions span diverse anatomical regions due to CT’s higher spatial resolution and different physical principles compared with PET. Indeed, in the BCR model, PET yielded 86 robust features versus only 7 from CT, most belonging to the low-frequency domain; these components are crucial for establishing a reliable training set in PET-based predictive modeling. Overall, PET imaging continues to demonstrate its suitability for radiomic analysis and predictive modeling, offering a more stable and informative feature set across heterogeneous anatomical contexts. This observation is consistent with findings by Urso et al., where PET-based radiomics outperformed CT in predicting pathological complete response in breast cancer, further underscoring the potential of PET as a robust foundation for radiomics-driven clinical applications [16].

Nonetheless, the clinical translation of radiomics faces hurdles. Many radiomic parameters—including first- and second-order features and complex wavelet- or texture-based descriptors—lack intuitive clinical meaning, hindering routine adoption [28]. Bridging this gap will require advanced analytics and a better contextualization of high-dimensional radiomic data within known biological frameworks [29]. An important caveat of any predictive model is the balance between false positives (FP) and false negatives (FN). In our context, an FP, i.e., classifying a patient as likely to respond when they will not, could lead to the continuation of ineffective immunotherapy, unnecessary exposure to immune-related adverse events, and the delay of potentially life-saving alternative treatments. By contrast, an FN—misclassifying a true responder as a non-responder—could deprive a patient of a highly active therapy, potentially shortening their progression-free survival. In future clinical implementation, thresholds will need to be tuned to minimize the more clinically detrimental error type, and decision algorithms could incorporate FP/FN costs (e.g., via cost-sensitive learning) to reflect the asymmetric consequences in advanced cSCC management.

Methodological standardization is also essential—covering lesion segmentation (absolute vs. relative SUV thresholds), voxel discretization (fixed-bin number vs. fixed-bin width), and feature aggregation in multifocal disease [30]. For instance, a 40% SUVmax threshold balances reproducibility and accuracy, while fixed-bin-width discretization enhances consistency across reconstruction settings [31,32]. Dimensionality reduction techniques such as principal component analysis can streamline feature sets and reduce redundancy, as shown in melanoma radiomics [15]. Aggregation strategies, e.g., volume-weighted averaging across the three largest lesions, can translate lesion-level features into patient-level predictors. In addition, harmonization techniques, such as ComBat, are essential in multicenter or multi-scanner studies, where variability in acquisition protocols or reconstruction parameters can introduce batch effects that compromise feature comparability [33]. However, we would like to clarify that our study was conducted using a single PET/CT scanner, with the same acquisition and reconstruction protocol across all patients. No differences in CT kernel, PET reconstruction settings or SUV normalization approaches were present in the dataset. Therefore, the implementation of a harmonization step was not necessary in this context, as no batch effects related to equipment or protocol variability were expected.

All radiomic features in our study were extracted using the open-source platform PyRadiomics, which follows the guidelines of the Image Biomarker Standardization Initiative (IBSI) to ensure reproducibility and standardized nomenclature. In the set of features extracted, both from the CT and PET images, the robust and nonredundant features were identified by application of the Mann–Whitney U-test and Spearman correlation analysis. While recognizing that some texture-based features may seem mathematically abstract, previous studies have demonstrated their importance in characterizing tumor heterogeneity, metabolism, and aggressiveness in oncological imaging [34].

In this regard, although it would have been interesting to compare original and wavelet-based texture features separately, in our cohort, the limited number of robust features retained after reduction precluded a meaningful subgroup analysis. Such stratification would have compromised statistical power and model stability. Future studies with larger cohorts may enable more detailed comparisons.

Our study has some limitations, including its retrospective, single-center design and the limited number of patients, which may restrict generalizability; prospective validation in independent cohorts is needed before clinical implementation. Additionally, in our study, lesions were identified based on [¹⁸F]FDG uptake, and only the primary cutaneous lesion was considered for each patient. In cases with multiple lesions, as reported in Section 2.3, the one with the highest SUVmax was defined as the index lesion and used for radiomic analysis. Thus, each patient contributed a single lesion to the dataset.

Importantly, no spatial normalization or scaling procedures were applied to the features dataset during preprocessing. This choice preserves regional and morphological heterogeneity at the feature level, allowing the machine learning algorithm to incorporate this variability into the predictive modeling process. In this context, anatomical heterogeneity is not considered as noise but as a potentially informative component of the feature space. Nonetheless, we agree that anatomical location could represent a limitation to the external generalizability of our ML models.

Although SMOTE is a useful technique for mitigating class imbalance, it can potentially introduce synthetic artifacts and reduce model generalizability. In particular, oversampling may lead to overfitting or create unrealistic feature patterns, especially in small datasets. To minimize such risks, we performed SMOTE with a minimum number of three nearest neighbors, thus limiting the generation of overly similar synthetic samples. This approach helps preserve the intrinsic structure of the original dataset and reduces the likelihood of introducing artificial bias during the model construction [35].

The principal limitation of our study is the lack of external validation. This might limit the assessment of model generalizability. The observed drop in performance from AUC 0.97 (training) to 0.88 (internal validation) in the BCR_PET model reflects a reduction in the predictive performance in internal evaluation. However, we would like to note that the reduced performance on the independent test set compared to the training set suggests that the test data were truly unseen to the model. This supports the methodological rigor of our internal validation approach [36]. Unfortunately, we did not have access to an external dataset and there is no public database of [¹⁸F]FDG PET scans for advanced cSCC, reflecting the rarity of this clinical entity. Nonetheless, external validation—preferably using independent, multicenter datasets—remains crucial to confirm the robustness and reproducibility of our predictive radiomic signatures.

In our cohort, lesion segmentation was performed using a semi-automated thresholding approach based on 40% of SUVmax within the lesion bounding box. This method was selected because it reflects current clinical practice and is widely adopted in oncologic PET imaging due to its operational simplicity and reproducibility [37]. However, it does not assess inter-observer variability, nor does it formally assess robustness to different reconstruction kernels, as recommended by IBSI guidelines.

Looking forward, deep learning-based segmentation and feature extraction could streamline radiomic workflows, reducing manual VOI delineation and enabling near-real-time application [38]. In this regard, Hayat et al. introduced a transformer-based attention framework that fuses CNN-extracted features with handcrafted radiomics via a channel–spatial attention module, demonstrating significantly improved feature relevance and classification accuracy in breast cancer imaging [39]. This work highlights how integrating attention mechanisms can further enhance the discriminative power of radiomic–deep learning hybrids.

Prospective trials should embed radiomic biomarker collection to validate predictive algorithms and assess real-world utility. Beyond guiding individual patient care, our predictive radiomic signatures could be prospectively incorporated as stratification factors or exploratory endpoints in future clinical trials of first-line PD-1 blockade or combination regimens in advanced cSCC. By identifying likely responders at baseline, trial designers could enrich for sensitive subpopulations—boosting statistical power and reducing sample size—while also exploring adaptive designs where non-responder signatures trigger early protocol amendments. Embedding our radiomic models alongside translational assays (e.g., IL-8, PD-1, and T cells’ dynamics) would create a multiparametric biomarker platform, accelerating the shift toward precision immuno-oncology in rare skin cancers. Concurrently, expanding public repositories to encompass less-common pathologies—such as cSCC—and leveraging open data platforms like The Cancer Imaging Archive will facilitate multi-institutional external validation and accelerate the adoption of radiomics in PET imaging [40].

5. Conclusions

In this study, we developed a PET-based radiomic model that predicts immunotherapy benefit with an internally validated AUC of 0.88 (82% accuracy), outperforming CT-based models (AUC 0.75, 67% accuracy). CT radiomics predicted tumor grade with an AUC of 0.80 (73% accuracy), while PET radiomics achieved an AUC of 0.78 (77% accuracy). These noninvasive baseline imaging signatures could guide treatment selection, identify high-grade tumors for closer monitoring, and support clinical decisions at therapy initiation. Prospective, multicenter studies and integration with clinical and laboratory data are needed for clinical adoption.