Integrated primary tumor and all metastatic lymph nodes radiomics from 18 F-FDG PET/CT for precise M1 subclassification in nasopharyngeal carcinoma

Abstract

Background

The false-negative rate of approximately 18% in fluorine-18 fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) can lead to understaging by misclassifying M1b disease as M1a of nasopharyngeal carcinoma (NPC), which may result in the inappropriate administration of intensified therapy. This study aims to develop a ¹⁸F-FDG PET/CT-based prediction tool to assist clinicians in diagnosing TNM-9 M1 staging and guiding treatment decisions for patients with synchronous NPC (smNPC).

Methods

Ninety-seven patients with squamous cell carcinoma smNPC underwent pretreatment ¹⁸F-FDG PET/C. Radiomic features were extracted from primary tumors and all metastatic lymph nodes using CT and PET data. Ten metabolic parameters of the primary tumors and most FDG-avid metastatic lymph node, measured through three methods, were also assessed. After feature selection, nine logistic regression-based radiomics models and one random forest model were developed. Model Performance was evaluated via 10 repeats×5-fold cross-validation mean the Area Under the Receiver Operating Characteristic Curve (AUC). The optimal prediction model was presented using a nomogram developed on the entire cohort and was validated with a calibration curve, decision curve analysis (DCA), and by determining its optimal cut-off value.

Results

The M1a and M1b patient groups showed no significant differences in baseline characteristics or metabolic parameters from the primary tumors and metastatic lymph nodes (p > 0.05). The combined primary tumors and metastatic lymph nodes radiomics model achieved the highest AUC of 0.907 in validation set. The final nomogram demonstrated excellent discrimination (AUC 0.963), with sensitivity of 98.1% and specificity of 86.0% at the optimal cutoff (0.439). Calibration was reliable (slope 0.853, intercept 0.132). DCA confirmed clinical utility, providing net benefit across a wide threshold range (0.135–0.700) with a maximum net benefit of 0.533 at a threshold probability of 0.050.

Conclusion

This study offers a complementary diagnostic tool for M1 staging of smNPC that may inform treatment decisions. The findings also provide metabolic evidence supporting the robustness of the TNM-9 stage IV classification system of NPC.

Trial registration

This retrospective study was in accordance with the Declaration of Helsinki and approved by the XXX.

Background

Nasopharyngeal carcinoma (NPC) shows a markedly higher incidence in endemic regions––especially China and Southeast Asia––but remains relatively rare in Western populations [1, 2]. Its global incidence and mortality are projected to increase substantially by 2040, particularly among Asian populations [1, 2]. Compared with other head and neck cancers, NPC exhibits a significantly higher propensity for distant metastasis (DM). Current epidemiological data indicate that 4–10% of new NPC cases present with synchronous metastatic disease (smNPC) [3, 4]. The most common sites of metastasis are the bones, lungs, and liver. Furthermore, the anatomically deep location of the nasopharynx often delays diagnosis, which largely contributes to this high rate of synchronous metastasis at presentation [5, 6]. The presence of DM indicates a substantially poorer prognosis. Despite aggressive multimodal therapy, patients with smNPC typically have a median overall survival of only 12–30 months [7, 8].

The clinical management of smNPC remains challenging due to its marked heterogeneity, primarily driven by various metastatic patterns and tumor burden (lesion number), which directly influence therapeutic efficacy and survival outcomes [9]. This recognition prompts ongoing refinements of the M1 classification [9–12], culminating in the 2025 American Joint Committee on Cancer (AJCC)/Union for International Cancer Control (UICC) TNM-9 system. This framework classifies all metastatic cases as stage IV and introduces subcategories IVA (M1a, ≤ 3 lesions) and IVB (M1b, > 3 lesions) to quantify metastatic burden for greater prognostic precision. Driven by revisions to the staging system, management strategies for M1-stage NPC have been refined. Current protocols advocate definitive therapy for M1a patients, contrasting with the continued recommendation of palliative care for those classified as M1b. By distinguishing patients according to disease extent, this stratification supports the selection of candidates for intensive, potentially curative regimens versus palliative approaches [13]. Hence, pretreatment mapping of DM, including their locations and total number, remains critical for clinical staging and treatment planning.

Fluorine-18 fluorodeoxyglucose positron emission tomography/computed tomography (¹⁸F-FDG PET/CT) is a well-established imaging modality for detecting DM, providing high sensitivity and specificity with a few false positives. Thus, it serves as a valuable alternative to conventional staging methods [14–17]. Nevertheless, a false-negative rate of approximately 18% poses a significant clinical risk. By failing to detect all metastatic lesions, it can lead to the misclassification of M1b as M1a disease, potentially resulting in inappropriately intensified treatment [16].

Consequently, studies have extensively investigated the molecular mechanisms underlying NPC metastasis to stratify the risk of DM [18–21]. Promising biomarkers, including the expression of deoxynucleotidyltransferase terminal-interacting protein 1 [19], integrin β3 levels in platelet-derived extracellular vesicles [20], and the circadian regulator ARNTL2 [21], are linked to metastatic propensity. In contrast, these prior investigations overlook the potential of metabolic parameters in predictive modeling. This gap is critical given that ¹⁸F-FDG PET/CT provides quantitative metabolic indices that capture tumor biology, such as primary tumor (T) aggressiveness, pretreatment staging, lymph node (LN) metastasis, and survival prognosis [22–25].

Advances in medical image analysis have enabled high-throughput extraction of quantitative features, driving the emergence of radiomics as a key methodology for transforming imaging data into clinical useful insights [26]. Reflecting this trend, a growing research has applied radiomics in NPC to enhance preoperative diagnostics, outcome prediction, and prognostic stratification [27].

The T and cervical LN metastases are recognized as significant factors associated with smNPC. This is supported by the identification of advanced N stage as an independent risk factor for DM [18]. Consistent with this observation, diagnostic models incorporating quantitative features from cervical metastatic LN show better predictive performance for smNPC than those relying solely on T characteristics [3].

Therefore, this study hypothesizes that Stage IVA and IVB smNPC exhibit significant radiomic and metabolic heterogeneity between T and metastatic LN. We aim to develop and validate a predictive model based on these features to accurately diagnose the M1 subcategory and guide treatment decisions.

Materials and methods

Patients

This retrospective study included patients with smNPC treated at our institution between April 2017 and October 2024. Inclusion criteria were: (1) age ≥ 18 years; (2) histologically confirmed squamous cell carcinoma through endoscopic biopsy; (3) available pretreatment ¹⁸F-FDG PET/CT and neck MRI; (4) TNM-9–based clinical staging. Exclusion criteria were: (1) incomplete visualization of the nasopharyngeal lesion on PET imaging and (2) a history of other malignant tumors.

Definition and verification of M1 substages (Reference Standard)

Temporal Definition of smNPC: DM was defined as disease identified at the initial diagnosis of NPC or within 3 months following the completion of primary radiotherapy.

Imaging-Based M1 Subclassification: As per the number-based criterion central to our study, patients with three or fewer radiologically confirmed metastatic lesions were classified as M1a, whereas those with more than three lesions were classified as M1b.

Verification and Follow-up Protocol: All metastatic lesions were confirmed either at initial presentation (via biopsy or characteristic imaging) or during follow-up. For lesions lacking immediate confirmation, patients underwent stringent imaging surveillance (e.g., PET/CT/MRI) every 3 months for ≥ 12 months. Lesions that progressed were definitively classified as DM, whereas those remaining stable or regressing were reclassified as benign or locoregional disease, ensuring diagnostic certainty.

PET/CT imaging

All patients underwent ¹⁸F-FDG PET/CT imaging on a Discovery 710 scanner (GE Medical Systems, Waukesha, WI, USA). After fasting for ≥ 6 h and confirming a blood glucose level below 11 mmol/L, each patient was administered an intravenous injection of ¹⁸F-FDG (0.1–0.2 mCi/kg). Patients then rested for 50–70 min in a quiet environment, during which they consumed 1000 mL of water and emptied their bladder before scanning. CT scans were performed from the vertex to the mid-thigh at 140 kV, using automated tube current modulation (Auto mA noise index, 28.5), with a rotation time of 0.8 s and a slice thickness of 3.75 mm. Subsequently, PET data were acquired across 6–7 bed positions with an emission time of 2 min per bed. Images were reconstructed using a fully 3D ordered-subset expectation maximization algorithm (VUE Point FX) with CT-based attenuation correction, employing the Sharp IR tool (24 subsets, 2 iterations, 192 × 192 matrix, 6.4 mm post-filter).

PET/CT-based metabolic features

Image analysis protocol

PET/CT images were independently analyzed by two blinded nuclear medicine physicians (with 5 and 8 years of experience, respectively) on an AW 4.6 workstation (GE Healthcare). First, volumes of interest (VOIs) for the T and the most metabolically active metastatic LN were semiautomatically delineated using the Advanced Workstation AW 4.6 (GE Healthcare) [28]. Then, manual slice-by-slice corrections were applied to these VOIs to exclude adjacent tissues such as brain or other non-lesional areas of high uptake.

Metabolic parameter extraction and definitions

Within the finalized VOIs, metabolic parameters were measured based on fixed Standardized Uptake Value (SUV) = 2.5, 40% SUV max and adaptive iterative (w = 0.5) thresholds [29].

SUVs: SUVmax (maximum SUV), SUVmean (mean SUV), and SUVpeak (peak SUV, mean within a small, fixed-size VOI centered on the hottest pixel) [30, 31].

Normalized Parameters: GNmax/GNmean (glucose-normalized SUVmax/SUVmean) [30, 32]; SULmax (lean body mass-normalized SUVmax) [33]; SUSmax (body surface area-normalized SUVmax) [34].

Volumetric Metrics: Metabolic Tumor Volume (MTV), defined as the total volume of tumor voxels [35]; Total Lesion Glycolysis (TLG), calculated as MTV × SUVmean [36]; and its glucose-normalized derivative, GNTLG (MTV × GNmean) [30].

PET/CT-based radiomics features

Volume of interest delineation and segmentation

Figure 1 illustrates the procedural workflow implemented in our study.

Fig. 1.

Flow chart of this study

Two radiologists (with 8 and 5 years of experience), blinded to clinical data, manually delineated the VOIs for T and all metastatic LN. Semiautomatic segmentation was conducted in 3D Slicer (v5.0.3; https://www.slicer.org/) using the PyRadiomics extension (http://www.radiomics.io/pyradiomics.html). The resulting VOIs for T and all metastatic LNs were then manually corrected slice by slice to exclude brain tissue.

Radiomics feature extraction

In total, 851 features were extracted from each VOI in CT and PET images using the PyRadiomics platform. These included first-order features (e.g., shape, volume, and histogram metrics) and second-order textural features derived from the gray-level co-occurrence matrix (GLCM), gray-level run-length matrix, gray-level dependence matrix, gray-level size zone matrix (GLSZM), and neighborhood gray-tone difference matrix (NGTDM) [37].

Statistical analyses

Software and modeling

All analyses were conducted in R (v4.5.1; http://www.r-project.org).

Radiomic features selection

Feature selection followed a two-stage approach: (1) preliminary screening with LASSO regression (glmnet) and (2) overfitting reduction using stepwise regression (MASS). Predictive models were developed via logistic regression (rms) and random forest (randomForest).

Model comparison

Model Performance was evaluated via 10 repeats×5-fold cross-validation mean the Area Under the Receiver Operating Characteristic Curve (AUC) (caret, pROC). To address the limited sample size, a repeated 5-fold cross-validation strategy was employed. The patient data were stratified and randomly partitioned into five folds. In each of the 10 iterations, the model was trained on four folds and validated on the remaining one, ensuring maximal data utilization and enhanced generalization. The final performance was reported as the mean AUC across all iterations.

Model demonstration

The optimal prediction model was presented as a nomogram developed on the entire cohort. Its performance was validated through a calibration curve, decision curve analysis (DCA), and determination of the optimal cut-off value (caret, rms, pROC).

Statistical tests

Inter-observer agreement was evaluated using intraclass correlation coefficients (ICC). Continuous variables were analyzed using Mann–Whitney U test (MTV, TLG, GNTLG) or the student’s t-test (demographics and standard PET parameters). Statistical significance was set as p < 0.05.

Results

Clinical and metabolic characteristics

The study cohort included 42 M1a and 55 M1b patients. Tables 1 and 2 detail clinical characteristics and metabolic parameters. Inter-observer agreement for all metabolic parameters was excellent (ICC > 0.75). Comparative analysis revealed no significant differences between the M1a and M1b groups in age, sex, height, weight, blood glucose level, or metabolic parameters derived from the T and the most active metastatic LN (all p > 0.05).

Table 1.

Patients’ characteristics

Characteristics	M1a (n = 42)	M1b (n = 55)	t/χ2	p
age	59.20 ± 10.09	52.29 ± 11.56	2.15	0.04*
Sex (Male/Female)	7/35	10/45	0.04	0.85
Height	166.81 ± 7.80	167.67 ± 7.87	−0.54	0.59
Weight	66.02 ± 13.29	66.51 ± 12.06	−0.19	0.85
blood glucose level (mmol/ml)	5.79 ± 0.90	5.75 ± 1.03	0.20	0.84
T stage
T1	1	4
T2	8	6
T3	9	17
T4	24	28
N stage
N1	7	13
N2	15	18
N3	20	24
Site of metastasis
Lung	6(14.28%)	7(12.72%)
Liver	4(9.52%)	2(3.64%)
Bone	29(69.05%)	20(36.36)
Lung and Bone	0	5(9.09%)
Lung and Liver	0	2(3.64%)
Liver and Bone	3(7.14%)	19(34.55%)

Table 2.

The comparison of metabolic parameters measured by the three different methods between the M1a and M1b groups

Metabolic parameters	Measurement methods	M1a	M1b	t/z	p
TMTV	SUV = 2.5	26.14(13.57,49.59)	25.21(15.34,48.85)	−0.13	0.9
	42%SUVmax	10.04(5.39,19.84)	9.35(4.78,16.69)	−0.33	0.74
	W = 0.5	14.17(8.23,23.04)	12.39(8.35,22.34)	−0.17	0.86
TSUVmax	SUV = 2.5	15.46 ± 6.07	15.20 ± 6.27	0.20	0.84
	42%SUVmax	15.46 ± 6.07	15.20 ± 6.27	0.20	0.84
	W = 0.5	15.46 ± 6.07	15.20 ± 6.26	0.20	0.84
TSUVmean	SUV = 2.5	5.94 ± 1.55	5.60 ± 1.25	1.19	0.24
	42%SUVmax	9.49 ± 3.87	9.10 ± 3.73	0.51	0.61
	W = 0.5	8.40 ± 3.47	7.89 ± 2.78	0.80	0.43
TTLG	SUV = 2.5	146.68(66.51,345.33)	149.72(79.98,248.10)	−0.12	0.91
	42%SUVmax	90.48(30.75,192.88)	79.62(49.16,138.93)	−0.60	0.55
	W = 0.5	107.37(49.82,232.41)	101.44(55.41,177.14)	−0.53	0.60
TGNmax	SUV = 2.5	14.98 ± 6.51	15.39 ± 6.40	−0.31	0.76
	42%SUVmax	14.98 ± 6.51	15.39 ± 6.40	−0.31	0.76
	W = 0.5	14.98 ± 6.51	15.39 ± 6.40	−0.31	0.76
TGNmean	SUV = 2.5	5.78 ± 1.85	5.71 ± 1.37	0.22	0.83
	42%SUVmax	9.18 ± 4.10	9.21 ± 3.82	−0.04	0.97
	W = 0.5	8.13 ± 3.69	7.99 ± 2.90	0.21	0.83
TGNTLG	SUV = 2.5	149.66(57.78,345.29)	159.31(81.28,243.75)	−0.06	0.95
	42%SUVmax	87.36(31.07,202.56)	83.66(48.55,148.40)	−0.38	0.70
	W = 0.5	102.14(42.84,242.61)	111.82(61.34,169.72)	−0.25	0.80
TSUVpeak	SUV = 2.5	12.62 ± 5.56	11.91 ± 4.90	0.67	0.50
	42%SUVmax	12.63 ± 5.55	11.91 ± 4.89	0.67	0.50
	W = 0.5	12.53 ± 5.64	11.90 ± 4.91	0.59	0.56
LNMTV	SUV = 2.5	12.43(5.81,18.78)	9.21(4.30,21.04)	−0.48	0.63
	42%SUVmax	5.32(2.19,7.45)	4.09(1.96,9.78)	−0.31	0.76
	W = 0.5	7.17(3.38,9.21)	5.43(3.00,11.91)	−0.25	0.8
LNSUVmax	SUV = 2.5	14.74 ± 5.15	14.77 ± 6.70	−0.03	0.98
	42%SUVmax	14.74 ± 5.15	14.78 ± 6.69	−0.03	0.97
	W = 0.5	14.74 ± 5.15	14.78 ± 6.69	−0.03	0.97
LNSUVmean	SUV = 2.5	6.11 ± 1.61	6.12 ± 2.06	−0.02	0.98
	42%SUVmax	9.22 ± 3.34	9.28 ± 4.45	−0.08	0.94
	W = 0.5	8.18 ± 2.97	8.19 ± 3.86	−0.02	0.98
LNTLG	SUV = 2.5	73.72(33.71,127.22)	55.88(22.18,127.77)	−0.5	0.62
	42%SUVmax	44.75(17.94,86.22)	32.94(14.95,87.46)	−0.5	0.62
	W = 0.5	54.75(25.44,103.17)	44.39(18.32,99.95)	−0.36	0.72
LNGNmax	SUV = 2.5	14.32 ± 5.64	14.94 ± 6.83	−0.48	0.63
	42%SUVmax	14.32 ± 5.64	14.94 ± 6.83	−0.48	0.63
	W = 0.5	14.32 ± 5.64	14.94 ± 6.83	−0.48	0.63
LNGNmean	SUV = 2.5	5.94 ± 1.88	6.19 ± 2.07	−0.62	0.54
	42%SUVmax	8.95 ± 3.66	9.38 ± 4.51	−0.5	0.62
	W = 0.5	7.95 ± 3.27	8.27 ± 3.86	−0.43	0.67
LNGNTLG	SUV = 2.5	75.25(26.94,129.64)	64.73(22.47,130.63)	−0.17	0.86
	42%SUVmax	42.52(15.54,85.38)	33.66(15.15,85.43)	−0.15	0.88
	W = 0.5	54.05(19.35,104.28)	48.14(18.55,98.01)	−0.07	0.94
LNSUVpeak	SUV = 2.5	11.38 ± 4.38	11.07 ± 5.33	0.30	0.76
	42%SUVmax	11.37 ± 4.39	11.05 ± 5.32	0.32	0.75
	W = 0.5	11.38 ± 4.36	11.05 ± 5.33	0.33	0.74
TSULmax	SUV = 2.5	12.24 ± 4.55	12.25 ± 5.29	−0.01	0.99
	42%SUVmax	12.24 ± 4.55	12.25 ± 5.29	−0.01	0.99
	W = 0.5	12.24 ± 4.55	12.25 ± 5.29	−0.01	0.99
LNSULmax	SUV = 2.5	11.78 ± 3.93	11.70 ± 5.13	0.08	0.93
	42%SUVmax	11.78 ± 3.93	11.70 ± 5.13	0.08	0.93
	W = 0.5	11.78 ± 3.93	11.70 ± 5.13	0.08	0.93
TSUSmax	SUV = 2.5	4.08 ± 1.53	4.11 ± 1.86	−0.07	0.95
	42%SUVmax	4.08 ± 1.53	4.11 ± 1.86	−0.07	0.95
	W = 0.5	4.08 ± 1.53	4.11 ± 1.86	−0.07	0.95
LNSUSmax	SUV = 2.5	3.92 ± 1.29	3.89 ± 1.77	0.08	0.94
	42%SUVmax	3.92 ± 1.29	3.89 ± 1.77	0.08	0.94
	W = 0.5	3.92 ± 1.29	3.89 ± 1.77	0.08	0.94

Radiomics models

Radiomic features selection

Radiomic features demonstrated excellent inter-observer agreement (all ICC > 0.75). A total of 851 features were extracted from the T and all metastatic LN on CT (TCT1–851, LNCT1–851) and PET (TPET1–851, LNPET1–851) images. A final logistic regression model was then developed with a two-step feature selection process using LASSO (Fig. 2) and stepwise regression (Table 3 lists selected features).

Fig. 2.

LASSO coefficient (λ) plots based on the TCT (a), TPET (b), TCT + PET (c), LNCT (d), LNPET (e), LNCT + PET (f), TCT + LNCT (g), TPET + LNPET (h), and TCT + LNCT + TPET + LNPET (i) data

Table 3.

The specific name of the radiomic features remained after feature screening and the AUC of the 9 logistic regression radiomics models

9 logistic regression radiomics models	after LASSO	after stepwise regression	AUC (training set)	AUC (validation set)
Model TCT	TCT198, TCT410, TCT418, TCT420, TCT494, TCT503, TCT517, TCT525, TCT703, TCT849.	TCT198, TCT410, TCT418, TCT494, TCT503.	0.817 ± 0.020 95% CI: (0.812, 0.823)	0.782 ± 0.088 95% CI: (0.757, 0.806)
Model TPET	TPET215, TPET246, TPET373, TPET432, TPET488, TPET518, TPET559, TPET581, TPET582, TPET747, TPET811.	TPET215, TPET373, TPET432, TPET488, TPET518, TPET559, TPET581, TPET747.	0.827 ± 0.027 95% CI: (0.819, 0.835)	0.753 ± 0.121 95% CI: (0.720, 0.787)
Model TCT + TPET	TCT198, TCT410, TCT420, TCT494, TCT517, TCT525, TCT703, TCT849, TPET373, TPET559, TPET582.	TCT198, TCT410, TCT517, TCT703, TPET373, TPET559, TPET582.	0.834 ± 0.023 95% CI: (0.828, 0.840)	0.773 ± 0.097 95% CI: (0.746, 0.800)
Model LNCT	LNCT8, LNCT186, LNCT188, LNCT389, LNCT402, LNCT438, LNCT559, LNCT645, LNCT675, LNCT740, LNCT747.	LNCT186, LNCT188, LNCT559, LNCT675, LNCT740.	0.820 ± 0.026 95% CI: (0.813, 0.828)	0.779 ± 0.119 95% CI: (0.746, 0.812)
Model LNPET	LNPET1, LNPET8, LNPET104, LNPET141, LNPET196, LNPET302, LNPET317, LNPET326, LNPET564, LNPET645, LNPET748	LNPET1, LNPET141, LNPET196, LNPET302, LNPET326, LNPET564, LNPET748.	0.838 ± 0.029 95% CI: (0.830, 0.846)	0.755 ± 0.122 95% CI: (0.722, 0.789)
Model LNCT + LNPET	LNCT8, LNCT186, LNCT438, LNCT559, LNCT645, LNCT675, LNCT740, LNPET116, LNPET141, LNPET302, LNPET748.	LNCT559, LNCT645, LNCT675, LNPET141, LNPET302,	0.838 ± 0.021 95% CI: (0.832, 0.844)	0.799 ± 0.100 95% CI: (0.771, 0.827)
Model TCT + LNCT	TCT198, TCT410, TCT420, TCT494, TCT517, TCT525, TCT703, TCT847, LNCT8, LNCT186, LNCT188, LNCT645, LNCT675, LNCT740	TCT198, TCT410, TCT517, TCT703, LNCT186, LNCT188, LNCT675, LNCT740	0.910 ± 0.017 95%CI: (0.905, 0.914)	0.857 ± 0.080 95% CI: (0.835, 0.879)
Model TPET + LNPET	TPET215, TPET246, TPET373, TPET401 TPET432, TPET488, TPET582, TPET747, LNPET8, LNPET52, LNPET141, LNPET196, LNPET302, LNPET645, LNPET748.	TPET373, TPET401 TPET432, TPET488, TPET747, LNPET8, LNPET141, LNPET196, LNPET302, LNPET748.	0.910 ± 0.019 95%CI: (0.905, 0.915)	0.853 ± 0.093 95% CI: (0.828, 0.879)
Model TCT + LNCT + TPET + LNPET	TCT53, TCT198, TCT410, TCT420, TCT494, TCT517, TCT525, TCT703, TCT849, LNCT8, LNCT186, LNCT188, LNCT475, LNCT645, LNCT675, LNCT740, TPET373, TPET559, TPET582, TPET747, LNPET52, LNPET141, LNPET196.	TCT198-wavelet-LLH-ngtdm-Complexity, TCT410-wavelet-HLL-glcm-Correlation, TCT420-wavelet-HLL-glcm-InverseVariance, TCT703-wavelet-HHH-glcm-MCC, LNCT186-wavelet-LLH-glszm- LargeAreaLowGrayLevelEmphasis, LNCT740-wavelet-HHH-glszm- GrayLevelVariance, TPET373-wavelet-LHH-glszm- LowGrayLevelZoneEmphasis, TPET559-wavelet-HLH-glszm- LowGrayLevelZoneEmphasis, TPET747-wavelet-HHH-glszm- SizeZoneNonUniformityNormalized, LNPET141-wavelet-LLH-glcm- InverseVariance, LNPET196-wavelet-LLH-ngtdm-Busyness.	0.974 ± 0.011 95%CI: (0.971, 0.977)	0.907 ± 0.077 95% CI: (0.886, 0.929)

Model comparison

The mean AUCs of logistic regression models based on the radiomics features of the T (training cohort: 0.817 ± 0.020, 0.827 ± 0.027, 0.834 ± 0.023; validation cohort: 0.782 ± 0.088, 0.753 ± 0.121, 0.773 ± 0.097) were comparable to those of models based on metastatic LN features (training cohort: 0.820 ± 0.026, 0.838 ± 0.029, 0.838 ± 0.021; validation cohort: 0.779 ± 0.119, 0.755 ± 0.122, 0.799 ± 0.100). The AUCs of the prediction models combining imaging features of the T and metastatic LN (training cohort: 0.910 ± 0.017, 0.910 ± 0.019, 0.974 ± 0.011; validation cohort: 0.779 ± 0.119, 0.853 ± 0.093, 0.907 ± 0.077) were higher than those of the models using either T or metastatic LN features alone. Among the nine models evaluated, the combined model (CT198-ngtdm, TCT410-glcm, TCT420-glcm, TCT703-glcm, LNCT186-glszm, LNCT740-glszm, TPET373-glszm, TPET559-glszm, TPET747-glszm, LNPET141-glcm and LNPET196-ngtdm) demonstrated the highest performance, with AUCs of 0.974 ± 0.011 and 0.907 ± 0.077 (the generalization gap was 0.057) in the training and validation cohorts, respectively (Table 3; Fig. 3).

Fig. 3.

10 repeats×5-fold cross-validation ROC curve of the training (blue line) and validation (orange line) cohorts for the model TCT (a), model TPET (b), model TCT + PET (c), model LNCT (d), model LNPET (e), model LNCT + PET (f), model TCT + LNCT (g), model TPET + LNPET (h), and model TCT + LNCT + TPET + LNPET (i)

In comparison, the random forest model identified the top 10 important features (Fig. 4) but showed modest performance (AUCs: 0.504 training, 0.660 testing) and excluded all metabolic parameters during construction.

Fig. 4.

Top 10 radiomics features in the random forest model (a). ROC curves for the training (red line) and validation (green line) cohorts (b)

Model demonstration

Development and Presentation of the Nomogram: The final multivariable logistic regression model, incorporating significant predictors from the radiomics feature sets, was developed using the entire cohort (n = 97). This model is visually presented as a nomogram (Fig. 5), allowing for the individualized estimation of the probability of requiring palliative care (Fig. 6).

Fig. 5.

Nomogram presentation (a) and validation (calibration curve (b), decision curve analysis (c), and ROC curves with optimal cut-off value (d) of the optimal prediction model developed on the full dataset

Fig. 6.

A 49-year-old male patient diagnosed with undifferentiated nasopharyngeal carcinoma by nasopharyngoscopy underwent PET/CT examination prior to treatment (a), which revealed a focal area of increased FDG uptake in the left sacrum (yellow arrow) without apparent abnormality in bone density. Nomogram model predicted an M1b risk value of nearly1 for this patient. Subsequent MRI demonstrated multiple abnormal signal foci in addition to the sacral lesion, involving both iliac bones (green arrows) and the left femur (not shown) (b) T1WI, (c) STIR, (d) DWI). These lesions resolved after treatment (e) T1WI, (f) STIR, (g) DWI). According to the TNM-9 staging system, the patient was classified as M1b

Model performance

The nomogram demonstrated excellent discriminative ability, achieving an AUC of 0.963. The optimal probability cut-off value, determined by maximizing Youden’s index, was 0.439. At this threshold, the model exhibited a sensitivity of 98.1% and a specificity of 86.0%. The calibration curve, assessed via bootstrap resampling (1000 repetitions), indicated good agreement between predicted probabilities and observed outcomes. The calibration slope was 0.853 and the intercept was 0.132 (ideal values: 1 and 0, respectively), suggesting the model’s predictions are generally reliable with a slight tendency towards overfitting and underestimation of risk in lower probability ranges (Fig. 5).

Clinical utility

The clinical net benefit of applying the prediction model was quantified using DCA across a spectrum of threshold probabilities (Fig. 5). The prevalence of the palliative care outcome in the cohort was 55.7%. The DCA revealed that the use of the nomogram to guide referral decisions provided a superior standardized net benefit compared to the strategies of “referring all patients to palliative care” or “referring all to active treatment” across a wide and clinically relevant threshold probability range from 0.135 to 0.700. The maximum standardized net benefit of 0.533 was observed at a threshold probability of 0.050.

Discussion

A key finding of this study is the absence of significant metabolic differences between the M1a and M1b subgroups, consistently observed across all three measurement methods. This consistency also appears in the feature-selection process, in which no metabolic parameters remain in the random forest model. Previous research has shown that measurement methodologies, especially thresholding techniques, can cause substantial variation in quantitative metrics like MTV and TLG [29]. Methodologies employed across existing studies often complicate meaningful comparison of their findings. Additionally, because the TNM-9 staging system is relatively new, its clinical use remains limited, and investigations assessing metabolic parameters within this framework are still scarce. Earlier research primarily utilized MTV and TLG values from T, metastatic LN, or DM to stratify risk among M1-stage patients and identify candidates likely to benefit from local radiotherapy [11]. In contrast, the TNM-9 staging system for NPC subclassifies stage IV disease solely based on the number of DM, independent of primary-tumor burden or nodal involvement. Our functional-imaging findings of no significant metabolic disparity between M1a and M1b groups support the biologic and clinical rationale for the TNM-9 stage IV classification update.

Predictive models for M1-stage stratification in NPC were developed in this study using radiomic features extracted from T and metastatic LN on PET and CT imaging. Across the nine logistic regression models constructed, those incorporating T features consistently were comparable to models relying solely on LN features. This finding contrasts with prior findings suggesting that LN characteristics possess greater diagnostic value for DM [3], including studies identifying advanced N stage as an independent risk factor [18]. A key distinction lies in the imaging modalities employed. Our predictive model is based on 18 F-FDG PET/CT, which provides complementary metabolic information. In contrast, the prior study utilized MRI-based features that primarily reflect anatomical and functional characteristics [3, 18]. This difference highlights the distinct informational basis of our approach. Furthermore, combined models integrating features from T and LN demonstrated better predictive performance than models based on a single-region model. The optimal model was a logistic regression based on 11 radiomic features: four from T-CT, two from LN-CT, three from T-PET, and two from LN-PET (TCT198-ngtdm, TCT410-glcm, TCT420-glcm, TCT703-glcm, LNCT186-glszm, LNCT740-glszm, TPET373-glszm, TPET559-glszm, TPET747-glszm, LNPET141-glcm, LNPET196-ngtdm).These findings collectively underscore that DM in NPC is influenced by the biological characteristics of the T and metastatic LN, supporting the integrative value of multiregion radiomic profiling for precise M-stage classification.

Our analysis identified several significant features from the NGTDM, GLCM, and GLSZM families, each reflecting tumor-texture heterogeneity and local intensity variations. This finding is consistent with that of prior radiomic studies across different cancer types. For example, in head and neck squamous cell carcinoma, NGTDM and GLSZM features are strongly prognostic for progression-free survival [38]. Similarly, in breast cancer, GLCM-based texture metrics from peritumoral adipose tissue on FDG PET/CT show potential for predicting chemotherapy response [39]. The recurrent importance of these feature classes across diverse cancers and imaging modalities corroborates their biological relevance. It also underscores their utility as robust imaging biomarkers in oncology.

In our comparative analysis of modeling techniques, multivariate logistic regression revealed substantially greater diagnostic performance than that of the random forest model. Specifically, the comprehensive logistic regression model achieved an AUC of 0.971, considerably exceeding that of the random forest model (AUC = 0.504). This finding is consistent with that of previous reports indicating that model performance, as measured through the AUC, is highly dependent on the feature selection strategy employed [40, 41]. The superior performance of logistic regression underscores its efficacy for this specific predictive task.

The primary objective of this study was to develop a complementary diagnostic tool for M-stage classification by predicting the likelihood of extensive metastasis (M1b) based on radiomic signatures from the T and metastatic LN. This study is motivated by a recognized clinical dilemma. PET/CT holds a paramount role in detecting DM. However, it is not infallible and has non-negligible rates of both false positives and false negatives. These inaccuracies can lead to staging errors and subsequent management uncertainties. Our model does not seek to replace PET/CT of whole body but to augment the diagnostic workflow by providing an independent, data-driven probability estimate derived from routinely acquired images. This “imaging biomarker” approach aims to refine risk stratification within the M1 category, potentially flagging patients who might benefit from more intensive scrutiny or confirming the confidence in the given PET/CT findings.

The exceptional discriminative performance of our nomogram (AUC = 0.963) strongly supports the biological and clinical plausibility of this complementary strategy. The high sensitivity (98.1%) at the optimal cut-off suggests the model is particularly effective at “ruling out” a low burden of metastasis, which could be valuable in supporting negative PET/CT findings. Conversely, the robust specificity (86.0%) indicates a good ability to “rule in” a high metastatic burden, potentially adding confidence to a positive PET/CT reading or highlighting cases where solitary PET-positive lesions might be suspicious for false positives.

The calibration metrics (slope = 0.853, intercept = 0.132) reveal that the model’s predicted probabilities are reliable for ranking patient risk but tend to be modestly conservative. This is acceptable and even prudent for a complementary tool designed for risk alert rather than definitive diagnosis. Most compelling is the decision curve analysis, which demonstrates the tool’s clinical utility across an impressively wide range of decision thresholds (0.135–0.70). This means that whether a clinician requires high certainty or is willing to act on a lower probability, integrating this model’s output into decision-making provides a net benefit over blanket strategies. The theoretical maximum benefit at a very low threshold (0.05) aligns with its role as a sensitive early-alert system.

To our knowledge, this is the first study to propose a radiomics model specifically for supplementing M-substage evaluation. By translating complex imaging data from T and metastatic LN into a clinically accessible nomogram, we offer a potential “second lens” to view metastasis risk. In practice, a high probability score from this model in a patient with equivocal or negative PET/CT findings could justify additional follow-up or alternative imaging, while a low score could reinforce confidence in a favorable M1a classification.

Our study has several limitations. First, it is a single-center, retrospective study with inherent selection biases. The high event prevalence (55.7%) and little sample size in our cohort may not reflect other settings, and external validation in prospective, multi-center cohorts is essential to confirm generalizability. Second, while we employed bootstrap internal validation and bias correction, the observed miscalibration underscores the need for model updating or recalibration in new populations. Third, the workflow currently requires manual segmentation and feature extraction, which may limit immediate point-of-care use; future work should explore automated pipelines and integration with clinical data for real-time prediction.

Conclusion

A robust radiomics-based model was developed and validated in this study to distinguish M1a from M1b stages in patients with smNPC. The integrative approach, which combines PET and CT radiomic features from T and all metastatic LN, demonstrates superior predictive performance and achieves satisfactory discrimination in the validation cohort. Notably, the absence of significant metabolic differences between M1 subgroups reinforces the biologic rationale of the TNM-9 classification, which defines stage IV disease solely based on metastatic burden. Furthermore, the dominant contribution of texture-based features (NGTDM, GLCM, and GLSZM) underscores the role of tumor heterogeneity in metastatic behavior. While further multicenter validation remains necessary, our findings demonstrate the potential of noninvasive radiomics for refining M-stage subclassification and supporting individualized treatment strategies, thereby advancing precision oncology in NPC.

Abbreviations

NPC

Nasopharyngeal carcinoma

DM

Distant metastasis

smNPC

Synchronous metastatic nasopharyngeal carcinoma

AJCC

American Joint Committee on Cancer

UICC

Union for International Cancer Control

¹⁸F-FDG PET/CT

Fluorine-18 fluorodeoxyglucose positron emission tomography/computed tomography

T

Primary tumor

LN

Lymph node

VOI

Volumes of interest

SUV

Standardized uptake value

SUVmax

Maximum SUV

SUVmean

Mean SUV

SUVpeak

Peak SUV

GNmax

Glucose-normalized SUVmax

GNmean

Glucose-normalized SUVmean

SULmax

Lean body mass-normalized SUVmax

SUSmax

Body surface area-normalized SUVmax

MTV

Metabolic tumor volume

TLG

Total lesion glycolysis

GNTLG

Glucose-normalized TLG

GLCM

Gray-level co-occurrence matrix

GLSZM

Gray-level size zone matrix

NGTDM

Neighborhood gray-tone difference matrix

AUC

Area under the curve

DCA

Decision curve analysis

ICC

Intraclass correlation coefficients

Authors’ contributions

YZ: study concept, design. YZ: analysis and interpretation of data. YZ, SX: data collection. YZ: drafting of the manuscript. YZ, YH: revision of the manuscript. All authors read and approved the final manuscript. All authors read and approved the final manuscript.

Funding

This study has received funding by the [talents program of Jiangsu Cancer Hospital #1] under Grant [number YC201801], [special project of clinical research on truth-seeking #2] under Grant [number ZL202210], [key project of Jiangsu Commission of Health #3] under Grant [number K2023021].

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethics approval and consent to participate

This retrospective study was in accordance with the Declaration of Helsinki and approved by the Institutional Ethics Review Boards of Jiangsu Cancer Hospital (protocol code ky-2024-132; approval date: December 12, 2024). The requirement for informed consent was waived because of the retrospective nature of the study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.