Longitudinal bilateral parotid delta-radiomic kinetics and variance signatures for prediction of late xerostomia following radiotherapy

Abstract

Background and purpose

Radiation-induced xerostomia remains a significant concern in head-and-neck cancer radiotherapy, inadequately predicted, with unidentified kinetic signatures of parotid injury. This study aimed to identify predictive metrics and reproducible bilateral parotid delta-radiomic kinetics and variance features for late xerostomia.

Materials and methods

This retrospective study included 131 HNC patients treated with VMAT-based radiochemotherapy. Weekly CBCT images (weeks 1–7) were collected (2020–2023). Twelve-month xerostomia was graded using CTCAE v5.0. IBSI-compliant bilateral parotid radiomics were extracted and converted to delta-radiomic kinetics. Patient-level data were split (training n: 91; testing n: 40), and feature selection combined univariate filtering with recursive feature elimination. Machine-learning models with stratified 5-fold cross-validation integrated clinical-dosimetric (CD), clinical-dosimetric + delta-radiomic (CDR), and clinical-dosimetric + delta-radiomic + kinetics (CDRK). Performance was evaluated using AUC (95% CI) and other metrics, with SHAP interpretability. Reliability (ICC, CCC, R²), multicollinearity, and bilateral asymmetry were assessed.

Results

The prevalence of 12-month grade ≥ 2 xerostomia was nearly identical between cohorts (76.9%, 77.5%; SMD: 0.014), with no significant difference observed. Predictive performance enhanced with feature incorporation and peaked mid-treatment. Clinical-dosimetric models demonstrated limited early discrimination (AUC 0.38–0.45). Delta-radiomics improved performance at week-3, while combined kinetic models achieved optimal predictive performance at week-4 (SVM-CDRK AUC 0.81, p: 0.002, sensitivity 100%, specificity 86%, Brier score 0.06, F1-score 91%). LASSO- and MLP-CDRK yielded AUCs of 0.75 (p: 0.009) and 0.73 (p: 0.011) and F1-score (86%; 79%). Texture-based GLCM features (AUC 0.79, p: 0.003) outperformed first-order and morphological descriptors. S_norm-, AUC_lc-, and variance-based kinetics showed high longitudinal reliability (ICC/CCC > 0.86; R² 0.78–0.83), revealing week-4 as a potentially informative treatment phase.

Conclusions

Longitudinal bilateral parotid delta-radiomic kinetics significantly enhance late xerostomia prediction, identify mid-treatment as the optimal risk-stratification window, and support adaptive radiotherapy and personalized toxicity mitigation.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12880-026-02573-z.

Introduction

Radiotherapy (RT) is a standard treatment for head-and-neck cancers (HNC); however, it frequently induces parotid gland injury, leading to late xerostomia and significantly diminishing patient quality of life. Despite advances in conformal techniques, including parotid-sparing image-guided RT and volumetric-modulated arc therapy (VMAT), a subset of patients still experiences clinically relevant salivary dysfunction, underscoring the need for biologically informed toxicity prediction beyond conventional dose-volume metrics [1–3]. Although the mean parotid dose (D_mean) is a validated predictor, dose-based normal tissue complication probability (NTCP) models do not fully capture inter-individual variability, suggesting that dynamic tissue remodeling during RT could provide additional prognostic insights [2, 4–6].

Radiomics has emerged as a promising approach to address this limitation by extracting high-dimensional quantitative features from longitudinally acquired standard medical imaging, enabling non-invasive quantification of tissue heterogeneity and spatial organization associated with treatment response and prognosis [5, 7–12]. Delta-radiomics extends this paradigm by characterizing longitudinal feature evolution and identifying treatment-induced tissue alterations before clinical toxicity manifests [1–3, 5, 9, 13–18].

Most current studies rely on static feature comparisons or unilateral assessments, without systematically modeling longitudinal kinetics, variance dynamics, or bilateral asymmetries throughout treatment [1–6, 12, 19–21]. It remains unclear whether early radiomic fluctuations reflect transient inflammatory effects or whether mid-treatment dynamics better reflect dose-driven parotid remodeling linked to long-term salivary dysfunction. Temporal kinetic descriptors, including normalized slope (S_norm), area under the longitudinal curve (AUC_lc), variance, and composite kinetic index (k_x), quantify the trajectory and instability of radiomic evolution, potentially identifying critical inflection points where biological divergence is maximal [9, 16].

This study aimed to (i) characterize the longitudinal bilateral parotid delta-radiomic kinetic and variance signatures, (ii) evaluate their reliability and temporal evolution, and (iii) assess their incremental value in predicting 12-month xerostomia. Clinical-dosimetric and delta-radiomic kinetic features were incorporated into complementary machine-learning frameworks (support vector machines (SVM), least absolute shrinkage and selection operator (LASSO), and multilayer perceptron (MLP)). To our knowledge, this is the first study to introduce bilateral longitudinal kinetic signatures within these predictive models, potentially enhancing mid-treatment risk stratification and supporting adaptive, personalized RT strategies.

Materials and methods

Patients

This retrospective longitudinal study included histologically confirmed HNC patients treated with VMAT between January 2020 and January 2023 at the Department of Radiotherapy and Oncology, Kuala Lumpur General Hospital, Malaysia. Of 173 eligible patients, 42 (24.3%) were excluded due to death during treatment, incomplete weekly CBCT, unilateral parotid irradiation, single-fraction regimens, or unknown primary tumor site. The final cohort comprised 131 patients (77.1% male, 22.9% female; age range 23–87 years) with complete weekly CBCT and documented 12-month xerostomia assessment. No patients in the included cohort exhibited clinically significant xerostomia before RT. Relevant clinical variables, including tumor site and stage, chemotherapy regimen, ECOG status, and prescribed dose were documented. However, baseline salivary function, xerogenic medication history, submandibular and oral cavity doses, and baseline parotid gland volume were not comprehensively assessed in the present analysis. The selection flowchart and demographics are summarized in Fig. 1; Table 1.

Fig. 1.

Flowchart of the patient selection procedure for this study

Table 1.

Demographics of patients

Characteristics	Cohort (N: 131) (%)
Gender
Male	101 (77.1)
Female	30 (22.9)
Age, years
Average (range: 23–87)	59.40 (11.9)
Tumor site
Nasopharyngeal	93 (70.9)
Tongue	9 (6.9)
Tonsillar	7 (5.3)
Glottis	6 (4.6)
Maxillary sinus	5 (3.8)
Nasal cavity	4 (3.1)
Oropharyngeal	3 (2.3)
Alveolar ridge	1 (0.8)
Buccal mucosa	1 (0.8)
Hard palate	1 (0.8)
Laryngeal	1 (0.8)
Smoking habits
Ex/current smoker	41 (31.3)
Non-smoker	90 (68.7)
Alcohol habits
Ex/current drinker	29 (22.1)
Non-drinker	102 (77.9)
Comorbidity
Yes	77 (58.8)
No	54 (41.2)
ECOG performance status
0–1	129 (98.5)
2–3	2 (1.5)
Tumor classification
T1	23 (17.6)
T2	32 (24.4)
T3	30 (22.9)
T4	46 (35.1)
Concurrent radiochemotherapy
Yes
Cisplatin	98 (74.8)
Carboplatin	9 (6.9)
No	24 (18.3)
Prescribed dose
70 Gy/35#/7-weeks	92 (70.3)
70 Gy/33#/6.5-weeks	6 (4.6)
66 Gy/33#/6.5-weeks	8 (6.1)
66 Gy/30#/6-weeks	15 (11.5)
60 Gy/30#/6-weeks	10 (7.6)
Xerostomia grades
None-to-mild	grade ≤ 1, n: 30 (22.90)
Moderate-to-severe	grade ≥ 2, n: 101 (77.09)
Body Mass Index
	≤ 20	≥ 20
Week-1_RT	10 (7.6)	121 (92.4)
Week-2_RT	15 (11.5)	116 (88.6)
Week-3_RT	15 (11.5)	116 (88.6)
Week-4_RT	17 (12.9)	114 (87.0)
Week-5_RT	22 (16.8)	109 (83.2)
Week-6_RT	28 (21.4)	103 (78.6)
Week-7_RT	36 (27.9)	95 (72.5)

Treatment characteristics

Patients were immobilized using a standard five-point thermoplastic shell. Prescribed doses ranged from 60 to 70 Gy (30–35 fractions) over 6–7 weeks. VMAT was delivered with the RapidArc system (Eclipse, Varian Medical Systems, USA) and on-board kilovoltage image guidance. Most patients received induction cisplatin (100 mg/m²) or carboplatin (400 mg/m²) every 3-weeks, followed by weekly concurrent radiochemotherapy. Twelve-month xerostomia was graded using CTCAE v5.0 and dichotomized (none-to-mild, grade ≤ 1; moderate-to-severe, grade ≥ 2). This study was conducted in accordance with the ethical principles of the Declaration of Helsinki [22] and approved by the Ministry of Health Malaysia and Universiti Kebangsaan Malaysia Research Ethics Review Committee (approval number: 23-02977-07E; UKM PPI/111/8/JEP-2023-817), which waived the requirement for informed consent.

Image acquisition and tumor delineation

Planning CT (pCT) scans were acquired supine using a 32-slice Aquilion™ LB scanner (Canon Medical Systems, Japan) (120 kVp, 300 mAs, 3 mm slice) with contrast enhancement (70-mL iodinated contrast; 50s delay; 512 × 512-matrix). CBCT images were captured during the first three fractions and weekly thereafter using the RapidArc system (100 kVp, 25 mAs, 2.5 mm slice), employing filtered back-projection. Image quality assurance (QA) followed AAPM TG-142/TG-179 guidelines using a CatPhan phantom [23–24]. Bilateral parotids were contoured on baseline pCT, propagated to weekly CBCT with deformable image registration (DIR), and manually verified by experienced radiation oncologists. The contralateral gland was operationally defined as the gland receiving the lower mean dose (D_mean), rather than explicit tumor laterality classification. Weekly CBCT acquisitions over the seven-week treatment are illustrated in Figure S1.

Feature extraction and delta-radiomics quantification

Radiomic features were extracted from segmented bilateral parotids (LT_PG/RT_PG) using LIFEx (v7.8.0; https://www.lifexsoft.org/) [25], following IBSI recommendations [7, 26]. Image preprocessing included voxel resampling (1 × 1 × 1 mm³), intensity normalization (-1000 to 3000 HU), and fixed-bin discretization (4, 8, 16, 32, 64). Delta-radiomics features were computed each treatment week as the relative change from baseline [27]:

1

where denotes the radiomic feature value at week n (n: 2, 3, 4, 5, 6, 7), and corresponds to the pre-treatment scan (week 1). These longitudinal delta-radiomic trajectories served as input for kinetic descriptor extraction. Detailed extraction parameters are provided in Additional File 1.

Feature selection, predictive modeling, and kinetic descriptors

The dataset was stratified at the patient level into training (n: 91, 69.5%) and testing (n: 40, 30.5%) cohorts, with comparable clinical characteristics, before any feature selection, resampling, or hyperparameter tuning. This ensured that no data from the same patient appeared in both training and testing. For week-specific models, the same patient-level assignment was maintained across all weekly time points. Dimensionality reduction and overfitting control were implemented using hybrid feature selection (univariate filtering and recursive feature elimination (RFE)) strictly within each training fold, ensuring that no information from the validation fold influenced feature selection. SMOTE resampling was applied only to the training subset within each cross-validation fold and never to test data. This prevented any leakage of synthetic information into held-out or validation data. Hyperparameters were optimized within a nested five-fold cross-validation framework using GridSearchCV and randomized search (up to 1000 iterations), with tuning performed exclusively on the training folds and model performance evaluated on the corresponding validation folds, which remained completely unseen throughout feature selection, resampling, and tuning, following TRIPOD guidelines [28]. Data-leakage prevention is detailed in Additional File 2.

Models were developed in Python (v7.3.2, scikit-learn) and integrated clinical-dosimetric, delta-radiomic, and kinetic features within SVM, LASSO, and MLP frameworks. Stepwise clinical-dosimetric (CD), clinical-dosimetric + delta-radiomics (CDR), and clinical-dosimetric + delta-radiomics + kinetic (CDRK) models were evaluated weekly using ROC-AUC (95% CI) and other metrics. The CDR model used delta-radiomic features derived at individual treatment weeks, whereas the CDRK model integrated longitudinal kinetic descriptors (S_norm, AUC_lc, variance, and k_x) derived from the complete temporal evolution of the delta-radiomic features. Learning curve analysis was performed to assess model convergence and potential overfitting by comparing training and testing performance across a progressively increasing seven-week VMAT treatment period. Model interpretability was computed exclusively on the held-out test set to avoid bias, using SHAP (SHapley Additive exPlanations), and bilateral asymmetry was calculated from RT_PG and LT_PG differences.

For each delta-radiomic trajectory, the following kinetic descriptors were summarized and computed once per patient over seven weeks (weeks 1–7) at 12-months post-RT, as previously described [29–30]:

1. Normalized slope (): Quantifies the overall temporal rate of change of delta-radiomic feature during RT.

2

3

• 2.Area under the longitudinal curve (): Summarizes the cumulative longitudinal burden of radiomic alteration across treatment.

4

5

• 3.Variance: Measures the temporal instability or fluctuation of the longitudinal delta-radiomic trajectory.

6

• 4.Composite kinetic index (): Captures trajectories that exhibit both sustained directional evolution and substantial temporal heterogeneity during treatment.

7

where , N, , , and represent the baseline feature value, total number of time points, time corresponding to feature , average time points, and mean delta-radiomic value across all weeks.

Reliability and multicollinearity assessment

Longitudinal reliability relative to baseline was evaluated using intraclass and concordance correlation coefficients (ICC-CCC) and explained variance (R²). Multicollinearity was assessed with the variance inflation factor (VIF > 10 unacceptable) (Additional File 3). Highly correlated delta-radiomic features (Spearman (|p| ≥ 0.9) were filtered, retaining the most predictive feature.

Statistical analysis and assessment of the study design

Bilateral parotid comparisons were performed using paired statistical tests (p < 0.05), and effect sizes were quantified using Cohen’s d. Predictive performance was evaluated on the independent test cohort using AUCs with 95% CIs. Statistical significance between correlated ROC curves was assessed using the paired DeLong test. Cohort balance was evaluated using standardized mean differences (SMDs) as described in [31], with SMD < 0.1 considered indicative of negligible imbalance. Study design quality and reporting transparency were assessed using the Radiomics Quality Score (RQS, www.radiomics.world) [32] (Table S1) and the TRIPOD statement recommendations (www.tripod-statement.org) [28]. The TRIPOD “Prediction Model Development and Validation” checklist and adherence extraction framework were applied to assess methodological completeness and reporting quality (Table S2).

Results

Patients’ characteristics

Within the 131 patients, 22.9% developed none-to-mild and 77.1% moderate-to-severe xerostomia at 12-months, with nasopharyngeal carcinoma predominating (70.9%). Cisplatin-based radiochemotherapy was the most common treatment (74.8%), and tumor stages were as follows: T1 (17.6%), T2 (24.4%), T3 (22.9%), and T4 (35.1%) (Table 1). The training and test cohorts had well-balanced baseline clinical, dosimetric, and outcome characteristics, with no statistically significant inter-cohort differences (p > 0.05) (Table 2). Mean age was comparable between the training and test cohorts (59.93 ± 11.55, 58.20 ± 12.99 years; p: 0.469, SMD: 0.141), and bilateral parotid gland dose distributions were similar, including D_mean_LT_PG (39.77 ± 13.63, 40.71 ± 11.15 Gy; p: 0.681; SMD: 0.075) and D_mean_RT_PG (39.07 ± 12.34, 43.46 ± 12.45 Gy; p: 0.066; SMD: 0.354). Importantly, the prevalence of 12-month grade ≥ 2 xerostomia was nearly identical between cohorts (76.9%, 77.5%; SMD: 0.014), confirming minimal cohort imbalance and supporting robust model development and validation.

Table 2.

Baseline characteristics of training and test cohorts with p-values and SMDs

Variable	Training cohort (n: 91)	Test cohort (n: 40)	-value	SMD
Clinical variables
Age, years	59.93 ± 11.55	58.20 ± 12.99	0.469	0.141
Gender
Male	60.00 ± 12.73	62.24 ± 11.84	0.395	0.180
Female	54.89 ± 11.03	55.41 ± 15.09	0.927	0.040
Smoking status (Yes)	0.45 ± 0.29	0.49 ± 0.38	0.329	0.189
Alcohol habits (Yes)	0.43 ± 0.24	0.38 ± 0.17	0.381	0.164
Comorbidities (Yes)	0.62 ± 0.49	0.53 ± 0.51	0.345	0.182
Tumor site
Nasopharyngeal	65 (71.4%)	28 (70.0%)	0.563	0.112
Non-nasopharyngeal	26 (28.6%)	12 (30.0%)	0.608	0.071
ECOG performance status	0.53 ± 0.36	0.48 ± 0.35	0.894	0.025
Tumor classification
T1–T2, n (%)	38 (41.8%)	17 (42.5%)	0.535	0.120
T3–T4, n (%)	53 (58.2%)	23 (57.5%)	0.379	0.161
Systematic therapy	0.84 ± 0.48	1.00 ± 0.51	0.085	0.335
BMI	24.29 ± 4.00	24.27 ± 3.89	0.978	0.005
Dosimetric variables
Prescribed dose	1.84 ± 1.39	1.77 ± 1.35	0.817	0.044
Dmean_LT_PG (Gy)	39.77 ± 13.63	40.71 ± 11.15	0.681	0.075
Dmean_RT_PG (Gy)	39.07 ± 12.34	43.46 ± 12.45	0.066	0.354
Outcome variable
12-month xerostomia (grade ≥ 2), n (%)	Positive = 70 (76.9%) Negative = 21 (23.1%)	Positive = 31 (77.5%) Negative = 9 (22.5%)	1.00	0.014

Continuous variables are presented as mean ± standard deviation (SD)

Dynamic model predictive performance

Predictive performance exhibited a distinct temporal and hierarchical pattern across model configurations (Table 3; Fig. 2a–d, Figure S2–S3). At weeks 1–2, CD models performed poorly (AUC 0.38; 95% CI: 0.37–0.47, 0.45; 95% CI: 0.44–0.52), with modest sensitivity (67%–86%), accuracy (55%–71%), precision (70%–75%), and F1-score (46%–67%), but limited specificity (31%–50%) and a high Brier score (0.28–0.45). Adding CDR yielded modest improvement, particularly in SVM models (AUC ≤ 0.47, F1-score: 53%–72%, Brier score: 0.19–0.29).

Table 3.

Comparison of predictive performance of different models for each week at the 12-month endpoint

Treatment week	Models	Predictive metrics	AUC [95% CI]	SEN (%)	SPE (%)	ACC (%)	PRE (%)	BS	F1-score (%)
W1	CD	LASSO	0.39 ± 0.11 [CI 0.33, 0.44] p: 0.087	68	38	55	72	0.45	51
		SVM	0.41 ± 0.12 [CI 0.36, 0.48] p: 0.086	71	43	63	74	0.39	57
		MLP	0.38 ± 0.10 [CI 0.37, 0.47] p: 0.088	67	32	68	70	0.43	46
	CDR	LASSO	0.40 ± 0.09 [CI 0.36, 0.45] p: 0.075	78	31	60	74	0.29	59
		SVM	0.43 ± 0.09 [CI 0.41, 0.50] p: 0.073	85	41	74	77	0.22	61
		MLP	0.39 ± 0.12 [CI 0.36, 0.48] p: 0.076	81	32	68	75	0.24	53
	CDRK	LASSO	0.50 ± 0.11 [CI 0.48, 0.59] p: 0.071	80	48	70	80	0.24	65
		SVM	0.52 ± 0.10 [CI 0.50, 0.60] p: 0.071	87	53	80	83	0.21	72
		MLP	0.49 ± 0.10 [CI 0.48, 0.58] p: 0.075	79	47	80	78	0.22	54
W2	CD	LASSO	0.41 ± 0.13 [CI 0.37, 0.50] p: 0.062	77	41	63	71	0.36	63
		SVM	0.45 ± 0.08 [CI 0.44, 0.52] p: 0.060	86	50	71	75	0.28	67
		MLP	0.41 ± 0.12 [CI 0.39, 0.51] p: 0.060	81	43	63	73	0.34	55
	CDR	LASSO	0.43 ± 0.12 [CI 0.42, 0.54] p: 0.057	81	47	71	77	0.26	69
		SVM	0.47 ± 0.10 [CI 0.46, 0.56] p: 0.051	88	54	77	80	0.19	72
		MLP	0.40 ± 0.11 [CI 0.36, 0.47] p: 0.062	80	47	70	78	0.22	61
	CDRK	LASSO	0.54 ± 0.08 [CI 0.52, 0.60] p: 0.047	82	53	74	81	0.23	77
		SVM	0.56 ± 0.08 [CI 0.54, 0.62] p: 0.038	89	57	78	83	0.18	84
		MLP	0.52 ± 0.10 [CI 0.49, 0.59] p: 0.045	85	49	75	80	0.19	67
W3	CD	LASSO	0.52 ± 0.11 [CI 0.51, 0.62] p: 0.037	81	47	76	75	0.24	64
		SVM	0.56 ± 0.10 [CI 0.53, 0.63] p: 0.031	87	55	77	77	0.24	69
		MLP	0.50 ± 0.11 [CI 0.47, 0.58] p: 0.035	81	43	78	76	0.30	62
	CDR	LASSO	0.59 ± 0.11 [CI 0.55, 0.66] p: 0.028	86	62	80	83	0.19	71
		SVM	0.62 ± 0.09 [CI 0.61, 0.70] p: 0.023	89	76	80	85	0.16	76
		MLP	0.55 ± 0.11 [CI 0.50, 0.61] p: 0.025	87	65	77	85	0.22	66
	CDRK	LASSO	0.64 ± 0.09 [CI 0.59, 0.68] p: 0.019	86	73	81	87	0.13	80
		SVM	0.69 ± 0.08 [CI 0.64, 0.72] p: 0.013	90	79	87	90	0.11	88
		MLP	0.67 ± 0.11 [CI 0.59, 0.70] p: 0.017	88	70	79	85	0.17	74
W4	CD	LASSO	0.61 ± 0.11 [CI 0.57, 0.68] p: 0.032	86	66	76	81	0.23	70
		SVM	0.63 ± 0.10 [CI 0.60, 0.70] p: 0.028	89	74	78	84	0.19	75
		MLP	0.61 ± 0.12 [CI 0.60, 0.72] p: 0.029	85	55	77	83	0.24	68
	CDR	LASSO	0.70 ± 0.11 [CI 0.66, 0.77] p: 0.019	88	70	81	84	0.19	79
		SVM	0.76 ± 0.08 [CI 0.72, 0.80] p: 0.015	89	77	84	89	0.13	82
		MLP	0.69 ± 0.09 [CI 0.65, 0.74] p: 0.017	86	68	79	85	0.22	75
	CDRK	LASSO	0.75 ± 0.10 [CI 0.70, 0.80] p: 0.009	97	78	87	87	0.10	86
		SVM	0.81 ± 0.07 [CI 0.76, 0.83] p: 0.002	100	86	94	91	0.06	91
		MLP	0.73 ± 0.11 [CI 0.66, 0.77] p: 0.011	93	74	85	85	0.11	79
W5	CD	LASSO	0.57 ± 0.10 [CI 0.55, 0.65] p: 0.036	80	51	73	75	0.18	61
		SVM	0.62 ± 0.09 [CI 0.59, 0.68] p: 0.031	84	59	78	78	0.17	67
		MLP	0.54 ± 0.11 [CI 0.51, 0.62] p: 0.037	77	52	68	72	0.24	59
	CDR	LASSO	0.60 ± 0.10 [CI 0.55, 0.65] p: 0.029	80	63	73	79	0.23	68
		SVM	0.63 ± 0.11 [CI 0.54, 0.65] p: 0.026	86	71	77	84	0.17	74
		MLP	0.60 ± 0.08 [CI 0.56, 0.64] p: 0.031	81	63	68	77	0.22	63
	CDRK	LASSO	0.67 ± 0.12 [CI 0.60, 0.72] p: 0.021	85	66	80	84	0.19	76
		SVM	0.69 ± 0.10 [CI 0.65, 0.75] p: 0.018	90	79	85	86	0.15	86
		MLP	0.66 ± 0.11 [CI 0.62, 0.73] p: 0.022	87	71	83	81	0.16	70
W6	CD	LASSO	0.52 ± 0.11 [CI 0.50, 0.61] p: 0.049	74	43	65	75	0.22	54
		SVM	0.52 ± 0.08 [CI 0.49, 0.57] p: 0.044	77	46	72	78	0.20	59
		MLP	0.49 ± 0.11 [CI 0.45, 0.56] p: 0.051	77	44	67	76	0.28	50
	CDR	LASSO	0.50 ± 0.07 [CI 0.47, 0.54] p: 0.045	80	51	74	80	0.21	67
		SVM	0.57 ± 0.09 [CI 0.51, 0.60] p: 0.041	80	55	75	82	0.17	72
		MLP	0.52 ± 0.11 [CI 0.45, 0.56] p: 0.041	79	48	71	77	0.27	59
	CDRK	LASSO	0.60 ± 0.11 [CI 0.58, 0.69] p: 0.034	84	56	77	83	0.18	73
		SVM	0.62 ± 0.11 [CI 0.56, 0.67] p: 0.030	89	60	82	84	0.15	81
		MLP	0.59 ± 0.13 [CI 0.55, 0.68] p: 0.036	86	53	77	80	0.22	64
W7	CD	LASSO	0.44 ± 0.12 [CI 0.38, 0.50] p: 0.065	65	40	63	71	0.29	49
		SVM	0.48 ± 0.12 [CI 0.45, 0.57] p: 0.059	80	43	70	74	0.22	56
		MLP	0.37 ± 0.12 [CI 0.34, 0.46] p: 0.061	72	41	67	73	0.34	48
	CDR	LASSO	0.46 ± 0.11 [CI 0.40, 0.51] p: 0.055	68	46	69	74	0.26	58
		SVM	0.52 ± 0.11 [CI 0.47, 0.57] p: 0.054	76	53	72	78	0.23	65
		MLP	0.47 ± 0.09 [CI 0.42, 0.51] p: 0.057	68	51	68	74	0.30	51
	CDRK	LASSO	0.55 ± 0.12 [CI 0.48, 0.60] p: 0.045	80	57	70	78	0.21	62
		SVM	0.58 ± 0.11 [CI 0.55, 0.66] p: 0.041	90	69	80	80	0.19	77
		MLP	0.53 ± 0.11 [CI 0.51, 0.61] p: 0.044	79	55	74	76	0.25	55

Note: AUC: area under the receiver operating characteristic curve; SEN: sensitivity; SPE: specificity; ACC: accuracy; PRE: precision; BS: Brier score; CD: clinical-dosimetric; CDR: clinical-dosimetric + delta-radiomics; CDRK: clinical-dosimetric + delta-radiomics + kinetic; W1: week-1; W2: week-2; W3: week-3; W4: week-4; W5: week-5; W6: week-6; W7: week-7

Fig. 2.

Predictive performance across model configurations depicting a clear temporal and hierarchical trend: (a) AUC (95% CI) of the most robust performance model (SVM-CD, SVM-CDR, SVM-CDRK), (b) performance trends of AUC across all models, (c) Brier scores of the optimal predictive model (SVM-CD, SVM-CDR, SVM-CDRK), (d) AUC (95% CI) discrimination of longitudinal delta-radiomic kinetics models at week-4 (SVM-CDRK, LASSO-CDRK, MLP-CDRK)

Performance improved substantially at week-3, with combined kinetic models (CDRK) outperforming CD and CDR. SVM-CDRK achieved AUC 0.69 (95% CI: 0.64–0.72, p: 0.017, Brier score: 0.11, F1-score: 88%). Peak performance occurred at week-4 across models, despite overlapping confidence intervals with adjacent weeks: SVM-CDRK (AUC 0.81; 95% CI: 0.76–0.83, p: 0.002, sensitivity: 100%, specificity: 86%, accuracy: 94%, precision: 91%, Brier score: 0.06, F1-score: 91%); LASSO-CDRK (AUC 0.75; 95% CI: 0.70–0.80, p: 0.009, sensitivity: 97%, specificity: 78%, accuracy: 87%, precision: 87%, Brier score: 0.10, F1-score: 86%); MLP-CDRK (AUC 0.73; 95% CI: 0.66–0.77, p: 0.011, sensitivity: 93%, specificity: 74%, accuracy: 85%, precision: 85%, Brier score: 0.11, F1-score: 79%). Repeated resampling supported the stability of this mid-treatment performance peak. This represents the strongest observed predictive performance, in which parotid dynamics inform late toxicity prediction.

Beyond week-5, model discrimination declined; however, CDR (AUC: 0.46–0.63, F1-score: 51%–74%) and CDRK (AUC: 0.53–0.69, F1-score: 55%–86%) consistently outperformed CD (AUC: 0.37–0.62, F1-score: 48%–67%). The SVM-CDRK model showed stable convergence, with training ROC-AUC gradually decreasing and validation ROC-AUC increasing and stabilizing as the training sample size increased (Figure S4). The narrowing gap between training and validation performance indicates improved generalization and controlled model complexity rather than overfitting. In comparison, LASSO-CDRK (Figure S5) showed a more persistent performance gap, while MLP-CDRK (Figure S6) exhibited lower and more fluctuating validation performance, consistent with underfitting. Among the models, SVM showed the most robust and stable performance, followed by LASSO, while MLP displayed greater variability. Overall, incorporating delta-radiomics and kinetic descriptors provided specific and incremental benefits, with weeks 3–4 representing the optimal window for late xerostomia risk-stratification.

Longitudinal bilateral parotid delta-radiomics

SHAP-based longitudinal analysis identified dynamic predictors of 12-month xerostomia across treatment weeks. Model discrimination enhanced from early treatment, peaked at mid-treatment (week-4), and declined modestly towards week-7 (Table 4). Among the clinical-dosimetric variables, comorbidities (AUC 0.743; 95% CI: 0.69–0.77, p: 0.004) and D_mean_RT_PG (AUC 0.751; 95% CI: 0.68–0.79, p: 0.006) demonstrated moderate and consistent predictive value at week-4. Gender was modest (AUC 0.647), and staging was inconsistent.

Table 4.

Performance statistics of longitudinal parotid radiomic biomarkers for predicting 12-month xerostomia

Feature	Treatment week
	W1	W2	W3	W4	W5	W6	W7
Clinical + Dosimetric
Comorbids	0.609 (0.57–0.73) p: 0.037	0.651 (0.55–0.63) p: 0.007	0.682 (0.66–0.81) p: 0.021	0.743 (0.69–0.77) p: 0.004	0.685 (0.61–0.74) p: 0.025	0.670 (0.65–0.76) p: 0.006	0.643 (0.60–0.72) p: 0.010
Dmean_RT_PG	0.631 (0.55–0.69) p: 0.027	0.669 (0.58–0.75) p: 0.036	0.677 (0.65–0.75) p: 0.009	0.751 (0.68–0.79) p: 0.006	0.700 (0.65–0.74) p: 0.012	0.601 (0.56–0.65) p: 0.00001	0.683 (0.59–0.71) p: 0.003
Gender	0.521 (0.48–0.61) p: 0.049	0.567 (0.45–0.60) p: 0.070	0.606 (0.51–0.64) p: 0.018	0.647 (0.58–0.68) p: 0.025	0.619 (0.56–0.70) p: 0.049	0.547 (0.53–0.67) p: 0.023	0.586 (0.56–0.72) p: 0.007
Staging	0.461 (0.41–0.55) p: 0.022	0.569 (0.53–0.66) p: 0.064	0.562 (0.49–0.61) p: 0.024	0.644 (0.57–0.72) p: 0.086	0.602 (0.59–0.68) p: 0.052	0.577 (0.55–0.70) p: 0.019	0.476 (0.42–0.530 p: 0.079
Radiomics
GLCM_LT_PG_64_Normalised Inverse Difference-(IBSI: NDRX)	0.617 (0.53–0.70) p: 0.029	0.629 (0.50–0.65) p: 0.011	0.693 (0.66–0.74) p: 0.003	0.758 (0.72–0.80) p: 0.001	0.707 (0.63–0.76) p: 0.023	0.642 (0.60–0.74) p: 0.003	0.699 (0.65–0.79) p: 0.012
INTENSITY-BASED_RT_PG_16_Intensity Skewness-(IBSI: KE2A)	0.627 (0.58–0.65) p: 0.003	0.642 (0.51–0.71) p: 0.043	0.651 (0.63–0.78) p: 0.005	0.731 (0.70–0.82) p: 0.001	0.685 (0.64–0.77) p: 0.006	0.676 (0.60–0.81) p: 0.022	0.634 (0.58–0.71) p: 0.002
INTENSITY-HISTOGRAM_RT_PG_16_Intensity-Histogram Kurtosis-(IBSI: C3I7)	0.648 (0.61–0.72) p: 0.023	0.670 (0.62–0.76) p: 0.035	0.655 (0.58–0.70) p: 0.014	0.711 (0.69–0.83) p: 0.012	0.659 (0.58–0.72) p: 0.026	0.689 (0.56–0.71) p: 0.031	0.546 (0.52–0.67) p: 0.023
GLCM_RT_PG_16_Correlation-(IBSI: NI2N)	0.664 (0.63–0.75) p: 0.003	0.687 (0.66–0.79) p: 0.007	0.700 (0.65–0.73) p: 0.001	0.787 (0.73–0.81) p: 0.003	0.731 (0.69–0.83) p: 0.008	0.679 (0.63–0.72) p: 0.00002	0.708 (0.68–0.79) p: 0.009
INTENSITY-HISTOGRAM_RT_PG_32_Minimum Histogram Gradient Grey-Level-(IBSI: RHQZ)	0.583 (0.54–0.68) p: 0.052	0.596 (0.55–0.67) p: 0.088	0.635 (0.60–0.70) p: 0.009	0.681 (0.61–0.77) p: 0.017	0.646 (0.60–0.75) p: 0.044	0.621 (0.58–0.72) p: 0.015	0.625 (0.60–0.69) p: 0.0003
GLCM_LT_PG_32_Inverse Variance-(IBSI: E8JP)	0.493 (0.45–0.61) p: 0.026	0.621 (0.59–0.70) p: 0.012	0.637 (0.60–0.71) p: 0.030	0.655 (0.61–0.75) p: 0.016	0.611 (0.57–0.70) p: 0.019	0.573 (0.52–0.67) p: 0.054	0.519 (0.44–0.59) p: 0.031
INTENSITY-HISTOGRAM_RT_PG_4_Minimum Histogram Gradient Grey-Level-(IBSI: RHQZ)	0.606 (0.57–0.68) p: 0.047	0.630 (0.60–0.76) p: 0.022	0.622 (0.54–0.67) p: 0.012	0.649 (0.59–0.73) p: 0.015	0.627 (0.57–0.72) p: 0.009	0.644 (0.53–0.67) p: 0.032	0.594 (0.53–0.66) p: 0.003
INTENSITY-HISTOGRAM_LT_PG_64_Minimum Histogram Gradient-(IBSI: VQB3)	0.534 (0.50–0.63) p: 0.033	0.548 (0.46–0.60) p: 0.009	0.630 (0.57–0.73) p: 0.027	0.709 (0.64–0.78) p: 0.022	0.644 (0.59–0.71) p: 0.015	0.629 (0.57–0.73) p: 0.020	0.557 (0.49–0.63) p: 0.055
MORPHOLOGICAL_RT_PG_32_Compactness2 -(IBSI: BQWJ)	0.527 (0.49–0.64) p: 0.027	0.533 (0.47–0.62) p: 0.070	0.612 (0.60–0.67) p: 0.001	0.635 (0.58–0.73) p: 0.029	0.620 (0.50–0.65) p: 0.052	0.577 (0.55–0.70) p: 0.024	0.529 (0.48–0.58) p: 0.015
INTENSITY-HISTOGRAM_RT_PG_4_Maximum Histogram Gradient-(IBSI:12CE)	0.604 (0.58–0.71) p: 0.041	0.638 (0.61–0.75) p: 0.023	0.647 (0.58–0.71) p: 0.019	0.695 (0.56–0.71) p: 0.027	0.627 (0.59–0.75) p: 0.051	0.655 (0.52–0.68) p: 0.045	0.535 (0.50–0.64) p: 0.025
GLCM_LT_PG_4_Cluster Shade-(IBSI:7NFM)	0.516 (0.44–0.60) p: 0.055	0.548 (0.50–0.64) p: 0.009	0.629 (0.55–0.68) p: 0.017	0.655 (0.59–0.73) p: 0.028	0.648 (0.61–0.77) p: 0.068	0.633 (0.56–0.70) p: 0.023	0.521 (0.48–0.56) p: 0.012
GLCM_LT_PG_4_Inverse Variance-(IBSI: E8JP)	0.455 (0.40–0.51) p: 0.029	0.585 (0.50–0.66) p: 0.057	0.601 (0.52–0.66) p: 0.033	0.652 (0.55–0.70) p: 0.037	0.611 (0.57–0.69) p: 0.010	0.469 (0.43–0.60) p: 0.068	0.581 (0.47–0.62) p: 0.013
GLCM_RT_PG_64_Inverse Variance-(IBSI: E8JP)	0.530 (0.50–0.61) p: 0.013	0.599 (0.55–0.71) p: 0.051	0.655 (0.60–0.71) p: 0.016	0.684 (0.55–0.73) p: 0.071	0.659 (0.58–0.75) p: 0.072	0.610 (0.56–0.73) p: 0.079	0.538 (0.45–0.62) p: 0.019
GLSZM_LT_PG_64_Large Zone Low-Grey-Level Emphasis-(IBSI: YH51)	0.549 (0.48–0.65) p: 0.070	0.556 (0.46–0.61) p: 0.067	0.634 (0.57–0.69) p: 0.025	0.641 (0.53–0.71) p: 0.088	0.570 (0.52–0.69) p: 0.031	0.486 (0.42–0.56) p: 0.065	0.636 (0.59–0.75) p: 0.022
INTENSITY-HISTOGRAM_LT_PG_64_Intensity-Histogram 25th Percentile-(IBSI: No)	0.583 (0.50–0.68) p: 0.033	0.609 (0.53–0.69) p: 0.045	0.628 (0.57–0.73) p: 0.031	0.645 (0.56–0.70) p: 0.088	0.639 (0.59–0.74) p: 0.009	0.627 (0.54–0.67) p: 0.043	0.624 (0.57–0.70) p: 0.002
INTENSITY-HISTOGRAM_RT_PG_16_Intensity-Histogram Mode-(IBSI: AMMC)	0.485 (0.42–0.57) p: 0.038	0.507 (0.48–0.63) p: 0.047	0.501 (0.45–0.57) p: 0.006	0.636 (0.52–0.70) p: 0.099	0.591 (0.56–0.71) p: 0.025	0.522 (0.45–0.59) p: 0.028	0.523 (0.49–0.67) p: 0.071

Note: Fixed bin sizes: 4, 8, 16, 32, 64

The GLCM_RT_PG_16_correlation-(IBSI: NI2N) achieved the highest performance (AUC 0.787; 95% CI: 0.73–0.81, p: 0.003), followed by GLCM_LT_PG_64_normalised inverse difference-(IBSI: NDRX) (AUC 0.758; 95% CI: 0.72–0.80, p: 0.001). Intensity-based skewness and kurtosis yielded AUCs of 0.71–0.73, whereas morphological features were weaker and inconsistent.

SHAP explainability revealed that longitudinal delta-radiomic texture features primarily influenced predictions over dosimetric parameters (Fig. 3a–b). GLCM_RT_PG_16_correlation (mean SHAP 0.87) suggested increased post-treatment spatial correlation and progressive textural homogenization. Complementary metrics like inverse variance and normalized inverse difference metrics (mean |SHAP| 0.28–0.46) supported a loss of microstructural heterogeneity. Intensity-histogram-based features, including kurtosis (0.56), skewness (0.24), and gradient metrics (0.39–0.48), contributed substantially. Although D_mean remained influential (mean |SHAP| 0.67) but comparable to radiomic features.

Fig. 3.

SHAP-based explainability and feature significance: (a) SHAP feature importance, (b) SHAP summary plots

Logistic regression coefficients confirmed concordant directionality, with GLCM_correlation showing the strongest positive association (β 1.14). Overall, the findings link late xerostomia risk to progressive parotid texture homogenization and intensity redistribution, supporting delta-radiomic kinetics as a biologically meaningful predictive approach.

Reliability and explanatory power of delta-radiomic kinetic descriptors

Longitudinal reliability analysis relative to baseline revealed variability across delta-radiomic features and kinetic descriptors (Table S3). Intensity-histogram kurtosis-(IBSI: C317) exhibited consistently excellent temporal stability (ICC 0.886–0.998; CCC 0.866–0.954) and strong explanatory power (R² 0.776–0.830), indicating stable temporal evolution. GLSZM_large zone low-grey-level emphasis-(IBSI: YH51) enhanced progressively, achieving near-perfect reliability by weeks 6–7 (ICC 0.986–0.999; CCC 0.966–0.999), alongside R² (0.999). Conversely, GLCM_correlation-(IBSI: N12N) and intensity skewness-(IBSI: KE2A) showed moderate, time-dependent longitudinal reliability (ICC 0.265–0.684; ICC 0.297–0.682) with limited explanatory power (R² ≤ 0.449, < 0.413). Morphological compactness displayed variable early stability, but improved in the mid-treatment phase (ICC 0.995).

Normalized slope analysis indicated stable high changes in kurtosis, gradual late increases in compactness and skewness, and dynamic texture evolution characterized by rising GLCM_correlation and early plateauing GLSZM metrics (Fig. 4). AUClc and variance peaked at week-4 (Figure S7–S8), followed by a partial decline, but remained above early values. k_x (Figure S9) also peaked mid-treatment, reflecting limited early discriminatory power.

Fig. 4.

Temporal dynamics of delta-radiomic kinetics normalized slope (S_norm) in bilateral parotid glands across treatment weeks

VIF analysis (Figures S10–S12) showed time-dependent increases in multicollinearity for local-intensity and first-order features, particularly for variance- and AUC_lc-based descriptors. In contrast, higher-order texture and morphological features remained stable (VIF < 5). Additionally, the correlations among longitudinal delta-radiomics features confirmed reproducible inter-feature relationships, with strong coupling between S_norm and AUC_lc (p: 0.77–0.71).

Bilateral parotid delta-radiomic kinetic asymmetries

Bilateral delta-radiomic analysis demonstrated significant differences in right-left variance (p < 1 × 10^− 9; Cohen’s d: − 0.110 to − 0.621) (Table S4). AUC_lc asymmetries became significant from week-4 (p ≤ 3.92 × 10^− 3; d: − 0.257 to − 0.356), reflecting higher kinetic activity mid- to late-treatment. Both S_norm and k_x exhibited minimal, non-significant asymmetries (p > 0.10; |d| ≤ 0.144). Peak asymmetries occurred between weeks 4–6, coinciding with the highest delta-radiomic activity. The RQS score was 14/36, reflecting moderate methodological quality.

Discussion

This study demonstrates that longitudinal bilateral parotid delta-radiomics, particularly when combined with kinetic descriptors, provides a biologically meaningful and clinically relevant framework for predicting 12-month xerostomia after radiotherapy. Clinical-dosimetric models alone showed limited early predictive performance, with only modest improvement by mid-treatment, consistent with previous NTCP frameworks that incompletely capture inter-individual glandular response [2, 4–6]. In contrast, integrating delta-radiomics significantly improved predictive performance, supporting emerging evidence that longitudinal radiomic signatures capture subclinical tissue alterations preceding visible gland dysfunction [3, 9, 12, 15].

Notably, incorporating kinetic descriptors yielded optimal performance at week-4 (SVM-CDRK), highlighting a potentially informative treatment phase. The superior performance of nonlinear SVM-based integration highlights the importance of modeling dynamic, interacting biological processes rather than relying solely on static dose metrics [3–5, 7, 9, 15, 30, 33]. This aligns with radiobiological understanding that salivary gland injury evolves progressively, reflecting acinar depletion and microvascular compromise [3, 9, 15].

During the first two treatment weeks, radiomic fluctuations likely reflected transient or partially reversible changes, limiting discriminatory power. By weeks 3–4, cumulative radiation exposure appears sufficient to induce measurable divergence in glandular microarchitecture, as indicated by rising AUC_lc, variance, and composite kinetic indices. These concurrent kinetic peaks suggest maximal biological separation between patients who will ultimately develop moderate-to-severe xerostomia and those who will not. Beyond week-5, predictive performance declined, possibly indicating structural homogenization as damage becomes established, thereby reducing the incremental prognostic signal.

Texture-based radiomic features, particularly GLCM and GLSZM descriptors, outperformed first-order and morphological metrics, underscoring the potential importance of microstructural alterations in late salivary dysfunction risk stratification [33]. SHAP analysis further identified multiscale texture features and D_mean as the most influential predictors, indicating complementary contributions from imaging phenotype and dosimetric burden. Increases in GLCM_correlation and related inverse-variance-type features may reflect progressive textural homogenization of the parotid gland during treatment. These findings align with prior reports of radiation-induced salivary gland injury [12, 34–37] and may reflect underlying processes such as acinar loss, fibrosis, or microvascular alteration. Early treatment weeks showed limited discriminatory performance, potentially reflecting subclinical or reversible tissue changes, whereas week-4 appears to capture more cumulative radiation effects, enabling clearer radiomic separation and improved risk stratification [7, 12, 16, 34–40]. Reliability analyses further supported the robustness and translational potential of the identified biomarkers. S_norm- and variance-based descriptors, especially intensity-histogram kurtosis and GLSZM features, exhibited high longitudinal reliability and explanatory power, reflecting stable temporal evolution. Although AUC_lc and integrated k_x showed greater variability, both converged at week-4, reinforcing this interval as a potentially informative treatment phase. Increasing multicollinearity among first-order intensity features likely reflects convergence of cumulative radiation-induced changes in global signal intensity and gland volume, as well as other underlying processes [33, 35, 37]. In contrast, higher-order texture (GLCM, GLSZM) and morphological features maintained acceptable collinearity, supporting their role as complementary descriptors of spatially heterogeneous structural remodeling rather than redundant intensity effects.

Bilateral analysis revealed asymmetric temporal variance patterns between the parotid glands, suggesting potential side-specific heterogeneity during treatment. These findings demonstrate the importance of independent gland assessment rather than an averaged bilateral evaluation; however, the observed right-left differences in variance and AUC_lc-based kinetics may partially reflect underlying dose heterogeneity and treatment geometry in asymmetrically irradiated patients [41].

Clinically, patients exhibiting high-risk delta-radiomic kinetic signatures at weeks 3–4 may benefit from adaptive replanning or salivary-sparing interventions before irreversible dysfunction develops. The combination of strong discrimination, calibration, and SHAP-based interpretability enhances clinical credibility and facilitates incorporation into adaptive RT workflows. Importantly, the week-4 window aligns with practical treatment timelines, allowing intervention while meaningful dose redistribution remains feasible.

Despite promising findings, several limitations should be acknowledged. Longitudinal CBCT-derived radiomic features remain susceptible to scatter artifacts, image noise, reconstruction variability, and HU instability, which can compromise feature robustness. In addition, DIR-based contour propagation across weekly CBCT acquisitions may introduce geometric uncertainties during treatment-related anatomical changes. Although all propagated parotid contours were manually reviewed and corrected by experienced radiation oncologists, residual registration inaccuracies may still influence temporal feature estimation. Standardized acquisition protocols, AAPM TG-142/TG-179 QA procedures, IBSI-compliant radiomics extraction, and ICC/CCC-based robustness filtering were implemented to reduce variability; however, these measures cannot fully eliminate CBCT-related uncertainties. Furthermore, the single-center design, class imbalance, moderate RQS (14/36, 38.9%), and absence of external validation may limit generalizability. Nevertheless, the biologically coherent week-4 bilateral parotid delta-radiomic kinetics observed in this study support their potential as non-invasive imaging biomarkers for predicting late xerostomia. Prospective multi-institutional validation with harmonized imaging and multimodal integration is required before clinical implementation in adaptive HNCRT.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

HNCRT

Head and-Neck Cancer Radiotherapy

VMAT

Volumetric Modulated Arc Therapy

CBCT

Cone-beam Computed Tomography

ICC

Intraclass Correlation Coefficients

CCC

Concordance Correlation Coefficients

TRIPOD

Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis

QA

Quality Assurance

IBSI

Image Biomarker Standardization Initiative

SMOTE

Synthetic Minority Oversampling Technique

S_norm

Normalized Slope

AUC_lc

Area-Under-the-Curve of Longitudinal Change

Var

Temporal Variance

k_x

Composite Kinetic Index

RT

Radiotherapy

CI

Confidence Interval

VIF

Variance Inflation Factor

LIFEx

Local Image Feature Extraction

GLCM

Grey-Level Co-occurrence Matrix

GLSZM

Grey-Level Size Zone Matrix

NGTDM

Neighboring Grey Tone Difference Matrix

GLRLM

Gray Level Run Length Matrix

Author contributions

D.O.S.: Conceptualization, Methodology, Software, Data curation, Validation, Writing-original draft preparation, Writing – Reviewing and Editing. M.I.: Conceptualization, Data curation, Methodology, Software, Validation, Writing – Reviewing and Editing. M.A.M.H.: Conceptualization, Validation, Writing – Reviewing and Editing. E.I.M.M.: Conceptualization, Visualization, Writing – Reviewing and Editing. H.A.M.: Conceptualization, Methodology, Validation, Visualization, Writing – Reviewing and Editing. N.Y.: Conceptualization, Methodology, Supervision, Software, Validation, Funding acquisition, Writing-original draft preparation, Writing – Reviewing and Editing.

Funding

This research was supported by the Premier Impact Fund 2.0 (DIP 2.0) [Grant number: DIP-2024-032].

Data availability

Clinical imaging data are not publicly available due to privacy and institutional policies and will be shared upon request to the corresponding author. All other data generated or analyzed during the study are included within this published article and its supplementary materials. Machine learning code is available upon request for academic and non-commercial use.

Declarations

Ethics approval and consent to participate

This study was conducted in accordance with the ethical principles of the Declaration of Helsinki and approved by the Ministry of Health Malaysia and Universiti Kebangsaan Malaysia Research Ethics Review Committee (approval number: 23-02977-07E; UKM PPI/111/8/JEP-2023-817), which waived the requirement for informed consent. All patient data were anonymized prior to analysis to ensure confidentiality and compliance with institutional and ethical guidelines.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.