+
Skip to main content
BMC is moving to Springer Nature Link. Visit this journal in its new home.
Download PDF

Composite quantitative structural magnetic resonance imaging-based risk scoring model for predicting radiation-induced temporal lobe necrosis in nasopharyngeal carcinoma: a novel risk stratification model

Radiation Oncology volume 20, Article number: 160 (2025) Cite this article

Abstract

Background

Radiation-induced temporal lobe necrosis (TLN) impairs long-term survival of patients with nasopharyngeal carcinoma (NPC) after radiotherapy (RT). We aimed to develop an early scoring model that integrats quantitative MRI indicators and clinical factors to enhance TLN risk stratification.

Methods

Longitudinal MRI scans acquired pre-RT and within 6 months post-RT in 439 patients with NPC (67 necrotic vs. 811 normal temporal lobes) included three-dimensional T1-weighted imaging for gray matter macrostructures and diffusion tensor imaging for white matter microstructures. Clinical and combined models were built using Cox regression, and their performances were compared to evaluate the incremental value of quantitative MRI biomarkers. A composite structural MRI-based risk score (CSS) was constructed for the TLN risk stratification. The incidence of TLN was predicted using a logistic dose-response model.

Results

Combining quantitative MRI biomarkers with clinical factors, such as age, diabetes, and TL radiation dose, significantly improved predictive accuracy and increased the C-index to 0.888 (P = 0.018). CSS effectively identified individuals at high risk for TLN; those with high CSS had a significantly higher TLN risk than those with low CSS (hazard ratio (HR) [95% confidence interval (CI)] = 3.07 [1.77–5.33], P < 0.001). Individuals with high CSS required a lower 50% tolerance dose for 5-year TLN (72.0 Gy) than those with low CSS (75.2 Gy).

Conclusions

Our CSS quantitatively characterized the longitudinal structural alterations in the temporal lobes pre- and post-RT. Integrating CSS with clinical and dosimetric parameters enables accurate TLN risk stratification and informs personalized management for patients with NPC.

Clinical trial number

Not applicable.

Background

Nasopharyngeal carcinoma (NPC) is a common malignant tumor of the head and neck with a high incidence in southern China [1]. Radiotherapy (RT) is the primary treatment for NPC and can result in temporal lobe injury (TLI) as a major neurological complication because the temporal lobes (TLs) are involved in the RT target volume [2]. TLI typically progresses in three stages: acute, early delayed, and late-delayed. Early TLI, which includes acute and early delayed injuries [3], usually occurs within 6 months post-RT and appears normal on MRI in patients with mild neuropsychiatric symptoms. In contrast, late delayed injury typically begins ≥ 6 months post-RT and is progressive and irreversible [3]. Late delayed TLI can eventually evolve into temporal lobe necrosis (TLN), a condition that is incurable and severely reduces the patients’ quality of life [4]. Therefore, the early identification and prevention are critical.

Advanced age, higher T-stage tumors, and TL radiation doses have been identified as predictors of TLN risk [5]. These factors form the cornerstone of TLN risk prediction; however, they are insufficient to elucidate inter-individual TLN risk heterogeneity fully [6]. Furthermore, these clinical factors do not adequately capture the early pathological changes or characteristics of the tissue microenvironment [7]. Imaging modalities offer critical complementary insights into TLN risk assessments. Recently, studies integrating MRI-based radiomics with clinical predictors have shown improved TLN prediction compared with conventional clinical models [8, 9]. Nonetheless, radiomic approaches have frequently employed manual delineation of the entire TL or its white matter (WM) as regions of interest (ROIs), hindering precise evaluation of the association between injury to specific TL subregions and TLN occurrence [8, 10]. In addition, the lack of biological interpretation of many radiomic features limits clinical translation [11]. Consequently, a precise, robust, and clinical translational model is essential for predicting TLN.

Recent advances in quantitative structural neuroimaging have demonstrated high sensitivity in detecting radiation-induced injuries to gray matter (GM) and WM within normal-appearing TL parenchyma, providing new insights for predicting TLN risk [12,13,14]. Beyond detecting and quantifying minor and barely perceptible injuries, these neuroimaging metrics also elucidate potential damage mechanisms after RT. For example, GM metrics, such as reductions in cortical thickness and temporal subfield volume, suggest cortical atrophy and neuronal loss in early phase TLN post-RT [12, 15] Concurrently, WM metrics, including fractional anisotropy (FA) and radial diffusivity (RD) within 6 months post-RT [3, 13], suggest demyelination and vascular damage in temporal lobe WM fibers, aligning with post-RT pathological damage [16]. Standardized neuroimaging pipelines, specifically three-dimensional T1-weighted imaging (3D-T1WI) and diffusion tensor imaging (DTI), are widely available and commonly used to ensure the reproducibility of quantitative GM and WM metrics [3, 17]. Therefore, these neuroimaging metrics provide quantitative, biologically informed, and reproducible measures for detecting early GM and WM injuries post-RT and hold promise for enhancing the early prediction of TLN risk. However, most current studies are confined to observational or cross-sectional designs or rely on unimodal metrics that assess GM or WM independently [3, 18], which restricts a comprehensive assessment of TLN risk.

Integrating multimodal imaging data to construct composite MRI scores can comprehensively characterize neuropathological changes in patients with NPC after RT by integrating macroscopic GM structural and WM microstructural features, which are crucial for fully assessing TLN risk. In fields such as demyelinating diseases and neuro-oncology, multiparametric MRI-based composite scores have improved disease stratification and treatment response evaluation by leveraging synergistic biomarker interactions [19, 20]. Despite these successes, multimodal integration for TLN prediction remains largely unexplored, thereby representing a critical research gap.

Therefore, we aimed to construct a composite structural MRI-based risk score (CSS) by integrating longitudinal GM-WM structural MRI data acquired before and within 6 months after RT. Our objective was to assess its efficacy in improving the predictive accuracy and risk stratification of TLN in patients with NPC.

Methods

Participants

The Institutional Review Board of the Sun Yat-sen University Cancer Center approved this prospective study (SZR2022-083). After explaining the study procedure, all participants provided written informed consent. This study enrolled 439 patients with NPC. Patients were staged according to the 7th Edition of the American Joint Committee on Cancer Tumor Node Metastasis (TNM) classification, ranging from T1N0M0 to T4N4M0.

The inclusion criteria were as follows:

(1) Newly diagnosed patients with NPC with pathological confirmation and no evidence of distant metastasis; (2) patients aged 18–70 years; (3) those eligible for standard RT and underwent neuroimaging MRI at baseline and within 6 months after RT; and (4) Han Chinese ethnicity and right-handedness.

The exclusion criteria were as follows:

(1) Conventional MRI follow-up not completed for 6 months post-RT; (2) presence of metastatic disease (M1) before treatment, other concurrent malignancies, or recurrence during treatment or within 6 months post-RT; (3) TLI with other abnormalities in bilateral TLs and middle cranial fossa; and (4) missing clinical data (such as radiation dosage).

The diagnostic criteria are summarized in the Supplementary Information (SI). The enrollment pathway is illustrated in Fig. 1A.

Fig. 1
figure 1

Study flowchart. (A) Patient enrolment: All data of enrolled patients with nasopharyngeal carcinoma were collected from the Sun Yat-sen University Cancer Center, and 878 temporal lobes were randomly divided into a testing set and a training set in a 7:3 ratio. (B) Workflow schematic: After data processing and feature extraction, pre-/post/delta RT MRI indicators were used for model construction, followed by TLN prediction and risk stratification

Full size image

The treatment, follow-up, and diagnosis of TLN

All the patients were treated with intensity-modulated radiotherapy (n = 434) or tomotherapy (n = 5). Dose-volume histogram (DVH) parameters, including the minimum absorbed dose covering x% of the temporal volume (Dx), the ratio of the TL volume receiving more than x Gy (Vx), the maximum dose covering x cm3 volume of the TL (Dxcc), the maximum dose (Dmax), the mean dose (Dmean), and the minimum dose (Dmin) received by the bilateral TLs were acquired for subsequent analysis. DVH parameters were rescaled for a treatment schedule of 2 Gy per fraction (EQD2 Gy) to ensure comparability [5]. The detailed RT treatment plans and conventional MRI follow-up are explained in SI. Neuroimaging MRI, comprising 3D-T1WI and DTI, was performed before RT and within 6 months post-RT, coinciding with routine MRI follow-up for patients with NPC after RT. The diagnostic criteria for TLN are described in the SI. Of the patients with NPC, 59 developed TLN, resulting in 67 necrotic TLs after post-RT follow-up (the median follow-up time was 29 months [interquartile range: 6–96 months]).

Image acquisition and processing

MRI images were acquired using a GE Discovery MR 750 3.0 T scanner (Discovery MR 750, GE Medical Systems, WI, USA) at the Department of Medical Imaging, Sun Yat-sen University Cancer Center. Conventional MRI, 3D-T1WI, and DTI were performed for all patients. The MRI scanning parameters are provided in the SI.

Figure 1B shows the workflow schematic. The 3D-T1WI and DTI data were preprocessed and used for feature extraction and prediction. The detailed image preprocessing procedure is provided in the SI. Key DTI preprocessing steps included eddy current and motion correction (with gradient direction rotation) to mitigate distortions, followed by diffusion tensor fitting. Subsequently, individual diffusion metric maps were registered non-linearly to the MNI standard space. Mean values for each diffusion metric were then extracted from the normalized maps defined by the JHU-ICBM-DTI-81 atlas. Finally, the structural morphological features of the TL cortical regions (thickness, volume, and area), as well as the diffusion features (FA, Axial Diffusivity [AD], RD, and Mean Diffusivity [MD]) were extracted to construct the prediction model. To illustrate the artifact correction process in the temporal lobe regions, image processing flowcharts of the two temporal lobe sections are presented in sFig. 1. Three cases (3/439 of the total sample) were excluded due to failure to meet quality standards. The specific exclusion criteria included: excessive motion artifacts (visible signal dropouts or gross anatomical distortion) (1 case), severe geometric distortion significantly affecting registration accuracy (1 case), and poor signal-to-noise ratio preventing reliable segmentation (1 case). All scans were independently evaluated by two experienced neuroradiologists. Any discrepancies in quality assessment were resolved through consensus discussion to ensure consistent application of exclusion criteria across all cases.

Features construction

Notably, 135 structural features were extracted from the TL using the Destrieux (aparc.a2009s) atlas [21]. Fifteen TL subregions were analyzed using three morphological metrics: cortical thickness, surface area, and gray matter volume (GMV). These metrics were acquired at baseline and 6 months post-RT, with delta metrics (post-RT minus baseline measurements) computed for longitudinal change quantification. Delta GMV was specifically adjusted for total intracranial volume as a covariate in the analysis [22]. For diffusion MRI (dMRI), 72 features were derived from six WM fiber ROIs defined by the Johns Hopkins University WM label atlas [23]. Four diffusion parameters—FA, AD, RD, and MD—were extracted per ROI at identical time points (baseline, 6-month post-RT, and delta values between scans). Overall, 207 neuroimaging features, including 135 sMRI and 72 dMRI features per unilateral TL, were extracted for each patient (sTable 1).

Eleven clinical factors were retrieved from the database: age, gender, T stage, N stage, TNM stage, smoking behavior, alcohol drinking, diabetes, hypertension and therapeutic regimen. The therapeutic regimens of the enrolled patients included RT only, concurrent chemoradiotherapy (CCR), and a combination of neoadjuvant/adjuvant chemotherapy and concurrent chemoradiotherapy (NCT + CCR). In addition, 45 DVH parameters, including Dmax, Dmin, Dmean, D0.1 cc to D20cc, V10 to V50, and D10 to D80, were obtained along with the eleven clinical factors for subsequent model construction.

Statistical analysis

The TLs were randomly split into a training set (n = 620) and a test set (n = 258) in a 7:3 ratio. Clinical (clinical variables only) and combined (clinical and imaging variables) models were constructed in the training set and validated in the test set to determine the incremental prognostic value of structural MRI parameters over the clinical variables. After univariate analyses, variables with P values < 0.1 were selected in the multivariate Cox proportional hazards model to estimate the likelihood of hazard ratios (HRs) and their 95% confidence intervals (CIs). The Harrell concordance index (C-Index) in each cohort was computed to compare the discriminatory abilities of different models for prognostic prediction. C-indices were compared in a pairwise fashion using U-statistics, which were computed using the rcorrp.cens function of the Hmisc package in R. Net gain was compared between the models using the net reclassification index (NRI).

For the combined model, based on the quantitative structural MRI indicators selected in the multivariate Cox model, we proposed a composite structural MRI-based risk score (CSS) to represent the patient’s risk level after considering all selected imaging features. CSS was generated using a linear predictor of the parametric part of the Cox model hazard function. The formula for calculating the CSS is as follows:

$$CSS{\mkern 1mu} = {\mkern 1mu} {\beta _1}{x_1} + {\beta _2}{x_2} + \ldots + {\beta _n}_{{x_n}}$$

Here, β1, β2, …, βn are the regression coefficients for each image covariate, and x1, x2, …, xn are the corresponding covariate values [24]. The CSS calculation code is available in the GitHub repository at https://github.com/Panjie19980514/CSS-calculation. This repository provides a script that allows users to input the values for the ROIs in structural and diffusion MRI data to compute the CSS.

A nomogram incorporating clinical factors and the generated comprehensive indicator, CSS, was developed in the training set and validated in the test set to provide clinicians with a quantitative rather than a qualitative approach to TLN risk assessment and enable personalized TLN risk assessment. Subsequently, all calibration curves for 3 and 5 years were plotted.

Individuals were categorized into low- or high-image risk groups according to the median CSS (CSS < P50 and CSS ≥ P50, respectively). The 5-year TLN-free probability was calculated using the Kaplan–Meier method and compared using the log-rank test. In addition, image risk subgroups were reclassified by deciles, and stratification analyses were conducted for the different clinical subgroups to examine the robustness of CSS (see SI for detailed subgroup analysis).

Dose-response relationships were quantified via logistic regression (fitting protocol in SI), estimating 50% tolerance doses for the 5-year TLN probability (TD50/5). All statistical tests were two-sided, and statistical significance was set at P < 0.05. All analyses were performed using R version 3.6.1.

Results

This prospective study included 439 patients (315 men and 124 women; mean age: 39.82 ± 9.59 years). The median follow-up and latency periods from RT to TLN were 29 and 25 months, respectively. Regarding therapeutic regimens, 161 and 260 patients received CCR and NCT + CCR, respectively. Eighteen patients received RT only. The median interval between the completion of RT and subsequent neuroimaging MRI acquisition was 3 months, with an interquartile range spanning from 1 to 6 months. The median routine MRI follow-up period was 29 months (Interquartile Range, 6–97 months), and 59 patients (13.4%) developed TLN. We enrolled 67 necrotic and 878 normal-appearing TLs from 439 patients to build the TLN prediction model. The 3-year and 5-year TLN-free survival rates were 93.1% and 89.7%, respectively. Detailed patients’ demographics and clinical characteristics are presented in Table 1.

Table 1 Baseline characteristics of the training set and testing set
Full size table

The combined model demonstrated superior discriminative performance over the clinical model in the training set (C-index: 0.893 vs. 0.853, P = 0.006) and the entire cohort (C-index: 0.888 vs. 0.860, P = 0.018). In contrast, the marginal improvement in the test set C-index was not statistically significant (C-index: 0.873 vs. 0.869, P = 0.806). Notably, comprehensive analyses of NRI improvements revealed statistically significant enhancements across all the evaluation sets (all P < 0.05). For 5-year predictions, the combined model achieved remarkable NRI improvements of 64.3% (95% CI [0.223–0.912]; P = 0.002) in the test set, confirming the critical value of structural MRI biomarkers in refining prognostic precision. Detailed model comparison results are presented in sTable 2.

The detailed feature selection process for the combined model can be found in the SI. Following feature selection, a total of 25 clinical and imaging variables were included in the initial multivariate Cox regression model. After backward selection, 12 variables were ultimately retained. Table 2 displays the detailed findings of the univariate and multivariate Cox analyses. Cox regression analyses of the combined model revealed that age at diagnosis, diabetes, TL radiation dose parameters (Dmin and D0.5cc), and specific MRI biomarkers were significantly associated with TLN risk. GM biomarkers included involved specific TL cortical subregions, including the baseline thickness of the parahippocampal gyrus (PHG), delta volume and delta area in the medial PHG, delta area in the lateral superior temporal gyrus (STG), and delta thickness in STG subregions (planum polare and planum temporale). WM biomarkers focused on WM tract integrity, specifically post-RT FA reduction in the superior longitudinal fasciculus (SLF) and fornix-cres/stria terminalis. A TL radiation dose of D0.5 cc (HR = 9.59, 95% CI: 5.11–17.99, P = 1.95 × 10− 12) was significantly associated with a higher risk of TLN.

Table 2 Univariate and multivariate Cox proportional regression model of TLN in NPCs
Full size table

We established a nomogram integrating CSS indicators and several clinical factors, including age, diabetes, Dmin, and D0.5 cc, to predict the probability of TLN-free survival comprehensively. The nomogram is illustrated in Fig. 2A. Forest plots shows the risk estimates for each MRI indicator that constituting CSS (Fig. 2B). Figure 2C shows the anatomical locations of each MRI index, including the cortical areas and WM fibers, as displayed on the left TL. Figure 2D schematically shows the follow-up outcomes of the two patients with NPC with similar clinical characteristics but different CSS values to better illustrate the role of CSS in TLN risk stratification. In addition, the 3-year and 5-year calibration curves showed favorable agreement between the nomogram predictions and the observed probabilities in the training and test sets of the TLN cohort (sFig. 2).

Fig. 2
figure 2

Development and validation of a comprehensive nomogram for TLN risk stratification in patients with NPC. (A) Nomogram illustrating the risk assessment. (B) Forest plots displaying MRI-derived risk estimates for CSS components. (C) Anatomical visualization of CSS-associated MRI indices (cortical regions and white matter tracts), as displayed in the left temporal lobe. (D) Follow-up outcomes of two patients with NPC with matched clinical profiles but divergent CSS values, highlighting CSS-driven risk stratification

Full size image

As shown in sTable 3, we observed no significant differences in TLN risk compared to the baseline group (P0-P10) across the lower 50th percentile deciles. However, a significant increase in risk was noted starting from the P50-P60 decile group (HR = 12.25, 95% CI = 1.58–94.95, P = 0.016). This trend continued with progressively higher risks in the P60-P70, P80-P90, and P90-P100 groups than in the baseline (all P < 0.05), with the P90-P100 group showing the most pronounced elevation (HR = 17.53, 95% CI = 2.30-133.31, P = 0.006). The significant Ptrend (P = 1.2e-05) supports the validity of the decile stratification. Based on these findings, we redefined the patient cohort into high- and low-risk groups using the 50th percentile as the cutoff (P0-P50 and P50-P100). Specifically, individuals with high CSS (CSS > P50) had a significantly higher TLN risk than those with low CSS (CSS ≤ P50) (training cohort: HR = 2.99, 95% CI = 1.55–5.74, P = 0.001; test cohort: HR = 3.53, 95% CI = 1.27–9.80, P = 0.016; total cohort: HR = 3.07, 95% CI = 1.77–5.33, P = 6.51e-05; Table 3).

Table 3 The results of image risk groups in multivariate Cox regression model in training cohort, testing cohort and total cohort of TLN population
Full size table

In the clinical subgroup analyses (age: < 44 vs. ≥ 44 years; diabetes: with vs. without; TL radiation dose: D0.5cc < 64 Gy vs. ≥ 64 Gy; regimens: RT only, CCR, NCT + CCR; T stage: T1-2, T3-4, T3, T4), the CSS variable demonstrated excellent risk stratification capability (sFig. 3, sFig. 4, sFig. 5). We further addressed temporal variability through a sensitivity analysis, which validated the 3–6 month post-RT window as the optimal timeframe for obtaining reliable CSS-based risk predictions (sFig. 6). Specifically, individuals who were older, had diabetes, received high radiation doses, or had advanced T stage (T3–4) disease exhibited an elevated TLN risk. The CSS further refined risk stratification within these subgroups. This was evidenced by significantly elevated TLN risk among high-CSS patients: in the older adult subgroup, the 5-year TLN rate was 86% in the high-CSS group versus 71% in the low-CSS group, reflecting a 15% absolute increase in risk; similarly, the high-dose radiation subgroup showed a 22% increase (85% vs. 63%), and the advanced T-stage subgroup a 23% increase (85% vs. 62%) (all P < 0.001). In the diabetes cohort, high-CSS patients had a 27% higher 3-year TLN risk than those with low CSS (P = 0.11). Across treatment regimen subgroups, high-CSS patients consistently exhibited an 11–13% increase in TLN risk (RT group: P = 0.16; CCR group: P = 0.004; NCT + CCR group: P = 0.007). Most notably, in the T4 subgroup, high-CSS patients had a 30% increase in 5-year TLN risk compared to low-CSS patients, resulting in a TLN-free survival rate of only 49% (P = 0.002). A supplementary analysis revealed significant differences in D0.5 cc across T stages (KW test, P = 1e-38). DMP results (sFig. 7) demonstrated the following: T1 vs. T2, P > 0.05; T2 vs. T3, P = 5.3e-7; T3 vs. T4, P = 2e-21.

To further explore the potential clinical application of CSS, we evaluated the individualized RT tolerance dose of TLs based on individual risk. In the overall dataset, the TD50/5 of D0.5 cc for developing a TLN was 73.7 Gy. By dividing the CSS into different subgroups, we observed that individuals with high CSS had a higher TLN risk than those with low CSS (Fig. 3A, B, and C). A consistent trend was observed when both groups received the same radiation dose (Fig. 3D). Specifically, the TD50/5 of D0.5 cc was 72.0 Gy for individuals in the bottom 50% of CSS and 75.2 Gy for those in the top 50% of CSS. More tolerance doses for several probability levels are listed in sTable 4.

Fig. 3
figure 3

Clinical utility of CSS in TLN risk stratification and dose tolerance. (A-C) TLN risk stratification based on CSS subgroups with elevated risk observed in the high-CSS cohorts. (D) Consistent CSS-dependent TLN risk trends under equivalent radiation doses

Full size image

Discussion

In this study, we developed a comprehensive predictive model for TLN in NPC by integrating longitudinal quantitative GM-WM structural MRI metrics with clinical covariates (age, diabetes status, and TL radiation dose). This model achieved superior predictive accuracy compared to the clinical model (C-index: 0.888 vs. 0.860, P = 0.018). The CSS enabled robust TLN risk stratification, with patients with high CSS showing a 3.07-fold higher TLN risk and requiring stricter dose limits (D0.5 cc: 72.0 Gy vs. 75.2 Gy) than those with low CSS. Our model provides an intuitive and accurate tool for predicting TLN risk, thus facilitating personalized management of patients with NPC following RT.

By comprehensively evaluating longitudinal GM-WM damage across TL subregions, our combined model effectively captures premorbid vulnerability, tracks ongoing injury progression, and reveals residual risk heterogeneity that remains unaddressed by models that rely solely on clinical and dosimetric factors. Furthermore, our model can identify vulnerable TL subregions, potentially supporting the development of targeted RT strategies to safeguard high-risk areas, thereby providing an advantage over the conventional clinical models. Moreover, although prior research has explored TLN prediction using neuroimaging parameters, these studies were predominantly cross-sectional or typically relied on features derived from a single static state (pre/post-RT) [12, 18]. In contrast, we adopted a longitudinal approach, utilizing neuroimaging information from pre- and post-RT phases and considering the evolution of features to comprehensively characterize early RT-induced injuries over time. In addition, in previous studies, most models relied on singular neuroimaging metrics, such as isolated GM or WM features, limiting their ability to fully capture the characteristics of TLI, thereby reducing the accuracy of TLN prediction and clinical utility [25, 26]. Our framework integrates longitudinal multimodal GM-WM neuroimaging signatures to address these shortcomings and offers a robust and accurate tool for early TLN prediction.

This findings of this study underscore the need for tighter dose constraints in high-risk individuals and support the validity of our risk stratification. The risk stratification capabilities of CSS were further validated through multiple subgroup analyses. Specifically, by distinguishing between individuals with low and high CSS, we achieved differentiated TLN risk assessments across high-risk subgroups. For instance, in older adults and high-dose radiation groups, individuals with high CSS exhibited a 15% and 22% increase in the 5-year TLN risk, respectively. Furthermore, across different treatment regimen subgroups, patients with high CSS consistently showed an elevated TLN risk, with increases ranging approximately from 11% to 13%. Most notably, patients with high CSS in the advanced T-stage cohort, particularly T4 NPC showed a 30% higher 5-year TLN risk with a TLN-free rate of only 49%, underscoring the need for intense monitoring or timely interventions to prevent TLN. From this perspective, the CSS framework overcomes the limitations of traditional models in personalized prediction and provides clinicians with an intuitive imaging-based quantitative tool for assessing TLN risk. The ability of CSS to identify high-risk patients early, aligns with the growing emphasis on using quantitative imaging biomarkers for risk-adapted management in oncology [20].

For individuals with are at a higher inherent risk of TLN, such as patients with T4 NPC or elderly patients, proactive risk-adapted RT planning can be implemented, which integrates hippocampal protection [27] and strict temporal lobe dose constraints [28]. For patients identified as having a high risk of TLN through the CSS framework, preventive interventions should be prioritized post-RT to avert the onset of injury. Pharmacological prophylaxis, particularly the prophylactic use of memantine [29], may help preserve neural integrity by countering the excitotoxic pathways implicated in early radiation damage. Systemic therapies with central nervous system (CNS) activity (e.g., EGFR/ALK inhibitors) [30] can also be leveraged to prevent neurotoxicity at its source. Exploratory approaches, such as hyperbaric oxygen therapy [31] or anti-inflammatory drugs need to be validated for their preventive effects. Moreover, close routine MRI follow-up should be conducted for individuals at high risk according to CSS. On the other hand, multidisciplinary collaboration can also guide the optimization of treatment plans before radiotherapy and the early implementation of protective strategies, shifting the focus from reactive management to preventive neuroprotection.

CSS employs longitudinal tracking of TL GM-WM variations in patients with NPC to elucidate TLN pathogenesis. Greater pre-RT cortical thickness in the PHG, along with a larger delta-RT PHG area difference and a smaller delta-RT volume difference, was associated with increased TLN risk. One possible explanation is that this pattern could potentially reflect heightened radiosensitivity due to preexisting neuroinflammation and microvascular fragility that may exacerbate radiation-induced damage [32]. Delta-RT cortical degeneration in the STG subregions, marked by reduced cortical thickness and surface area, was associated with increases TLN risk, which may reflect radiation-induced neurogliovascular damage [18, 26]. Post-RT FA elevation in the fornix was associated with increased TLN risk. Given its proximity to radiation fields, we interpret that this increase in FA could suggest processes such as maladaptive axonal sprouting that could potentially disrupt connectivity [33]. Conversely, increased FA in the SLF distant from high-dose radiation was associated with reduced TLN risk, which could be indicative of compensatory myelination [34]. Collectively, CSS provides critical insight into the structural changes associated with TLN, while also generating testable hypotheses regarding the underlying biological mechanisms.

This study has several limitations. First, this study had a single-center design. Although rigorous sensitivity analyses were conducted (including randomized sample partitioning and stratified analyses across clinical subgroups), multicenter validation remains crucial for clinical translation. Second, CSS involving the TL subregions and imaging metrics lacked histopathological validation. Further animal studies are required to correlate these biomarkers with post-RT neuropathological changes. Third, magnetic susceptibility artifacts may affect the DTI-derived parameters of the temporal lobes, particularly near the mastoid and sphenoid sinus interfaces. Despite our efforts in quality control, its complete elimination remains. Future studies using advanced techniques could address these issues [35, 36]. Fourth, the CSS framework requires both pre- and post-RT metrics for risk stratification, which limits its use for pre-RT dose adjustments. Future work will explore a pre-RT model incorporating baseline imaging, clinical factors, and possibly patient-specific genetic factors to predict TLN risk and guide radiotherapy planning. Fifth, subgroup analysis reveals that CSS maintains robust predictive capacity during the 3–6 month post-RT window in 664 TLs, whereas immediate post-RT phase exhibits limited predictive value in 190 TLs, identifying the 3–6 month period as the optimal timeframe for predictive evaluation. However, sample imbalance may influence results, necessitating further validation through larger-scale studies with balanced cohorts. Finally, implementing our CSS model requires additional dMRI and sMRI sequences, resulting in a total increased scan time of about 10 min and potential limitations in resource-limited regions with high NPC prevalence.Nevertheless, it offers considerable clinical value by enabling early identification of high-risk patients for timely intervention.

Conclusion

This study provides evidence that integrating longitudinal GM-WM neuroimaging metrics with clinical factors significantly enhances the accuracy of TLN risk prediction compared to conventional clinical models. We developed a sophisticated composite structural MRI-based indicator, which quantifies risk scores for patients with NPC following RT with high interpretability. This tool empowers clinicians to deliver individualized and timely interventions tailored to heterogeneous TLN risk profiles of diverse patient cohorts, thereby enhancing patients’ quality of life.

Data availability

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

References

  1. Torre LA, Bray F, Siegel RL, Ferlay J, Lortet-Tieulent J, Jemal A. Global cancer statistics, 2012. CA Cancer J Clin. 2015;65:87–108.

    PubMed  Google Scholar 

  2. Lam TC, Wong FC, Leung TW, Ng SH, Tung SY. Clinical outcomes of 174 nasopharyngeal carcinoma patients with radiation-induced temporal lobe necrosis. Int J Radiat Oncol Biol Phys. 2012;82:e57–65.

    Article  PubMed  Google Scholar 

  3. Ding Z, Zhang H, Lv X-F, Xie F, Liu L, Qiu S, et al. Radiation-induced brain structural and functional abnormalities in presymptomatic phase and outcome prediction. Hum Brain Mapp. 2018;39:407–27.

    Article  PubMed  Google Scholar 

  4. Zheng Z, Wang B, Zhao Q, Zhang Y, Wei J, Meng L, et al. Research progress on mechanism and imaging of temporal lobe injury induced by radiotherapy for head and neck cancer. Eur Radiol. 2022;32:319–30.

    Article  PubMed  Google Scholar 

  5. Wen D-W, Lin L, Mao Y-P, Chen C-Y, Chen F-P, Wu C-F, et al. Normal tissue complication probability (NTCP) models for predicting Temporal lobe injury after intensity-modulated radiotherapy in nasopharyngeal carcinoma: A large registry-based retrospective study from China. Radiother Oncol. 2021;157:99–105.

    Article  PubMed  Google Scholar 

  6. Zhou X, Liu P, Wang X. Temporal lobe necrosis following radiotherapy in nasopharyngeal carcinoma: New insight into the management. Front Oncol. 2021;10:593487. Available from: https://doi.org/10.3389/fonc.2020.593487/fullhttps://www.frontiersin.org/journals/oncology/articles/

  7. Li P-J, Chen Y-Y. Narrative review: Temporal lobe necrosis after radiotherapy for nasopharyngeal carcinoma. Ann Nasopharynx Cancer. 2020;4(0):2616–4191. Available from: https://anpc.amegroups.org/article/view/5718

  8. Bao D, Zhao Y, Li L, Lin M, Zhu Z, Yuan M, et al. A MRI-based radiomics model predicting radiation-induced temporal lobe injury in nasopharyngeal carcinoma. Eur Radiol. 2022;32:6910–21.

    Article  PubMed  Google Scholar 

  9. Hou J, Li H, Zeng B, Pang P, Ai Z, Li F, et al. MRI-based radiomics nomogram for predicting temporal lobe injury after radiotherapy in nasopharyngeal carcinoma. Eur Radiol. 2022;32:1106–14.

    Article  PubMed  Google Scholar 

  10. Zhang B, Lian Z, Zhong L, Zhang X, Dong Y, Chen Q, et al. Machine-learning based MRI radiomics models for early detection of radiation-induced brain injury in nasopharyngeal carcinoma. BMC Cancer. 2020;20:502.

    Article  PubMed  PubMed Central  Google Scholar 

  11. Yip SS, Aerts HJ. Applications and limitations of radiomics. Phys Med Biol. 2016;61:R150–66.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  12. Zhang YM, Chen MN, Yi XP, Li L, Gao JM, Zhang JL, et al. Cortical surface area rather than cortical thickness potentially differentiates radiation encephalopathy at early stage in patients with nasopharyngeal carcinoma. Front Neurosci. 2018;12:599.

    Article  PubMed  PubMed Central  Google Scholar 

  13. Nagesh V, Tsien CI, Chenevert TL, Ross BD, Lawrence TS, Junick L, et al. Radiation-induced changes in normal-appearing white matter in patients with cerebral tumors: a diffusion tensor imaging study. Int J Radiat Oncol Biol Phys. 2008;70:1002–10.

    Article  PubMed  PubMed Central  Google Scholar 

  14. Lin J, Lv X, Niu M, Liu L, Chen J, Xie F, et al. Radiation-induced abnormal cortical thickness in patients with nasopharyngeal carcinoma after radiotherapy. Neuroimage Clin. 2017;14:610–21.

    Article  PubMed  PubMed Central  Google Scholar 

  15. Nagtegaal SHJ, David S, van der Boog ATJ, Leemans A, Verhoeff JJC. Changes in cortical thickness and volume after cranial radiation treatment: a systematic review. Radiother Oncol. 2019;135:33–42.

    Article  PubMed  Google Scholar 

  16. Oi S, Kokunai T, Ijichi A, Matsumoto S, Raimondi AJ. Radiation-induced brain damage in children–histological analysis of sequential tissue changes in 34 autopsy cases. Neurol Med Chir (Tokyo). 1990;30:36–42.

    Article  CAS  PubMed  Google Scholar 

  17. Guo Z, Han L, Yang Y, He H, Li J, Chen H, et al. Longitudinal brain structural alterations in patients with nasopharyngeal carcinoma early after radiotherapy. Neuroimage Clin. 2018;19:252–9.

    Article  PubMed  PubMed Central  Google Scholar 

  18. Lin X, Li Z, Chen S, Yang Y, He H, Lv X et al. Divergent white matter changes in patients with nasopharyngeal carcinoma post-radiotherapy with different outcomes: a potential biomarker for prediction of radiation necrosis. Eur Radiol. 2022;32(10):7036–47.

  19. Poonawalla AH, Datta S, Juneja V, Nelson F, Wolinsky JS, Cutter G, et al. Composite MRI scores improve correlation with EDSS in multiple sclerosis. Mult Scler. 2010;16:1117–25.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  20. Kang KM, Song J, Choi Y, Park C, Park JE, Kim HS, et al. MRI scoring systems for predicting isocitrate dehydrogenase mutation and chromosome 1p/19q codeletion in adult-type diffuse glioma lacking contrast enhancement. Radiology. 2024;311:e233120.

    Article  PubMed  Google Scholar 

  21. Destrieux C, Fischl B, Dale A, Halgren E. Automatic parcellation of human cortical gyri and sulci using standard anatomical nomenclature. NeuroImage. 2010;53:1–15.

    Article  PubMed  Google Scholar 

  22. Wu J, Gao M, Piao R, Feng N, Geng B, Liu P. Magnetic resonance imaging-based structural covariance changes of the striatum in lifelong premature ejaculation patients. J Magn Reson Imaging. 2022;55:443–50.

    Article  PubMed  Google Scholar 

  23. Hua K, Zhang J, Wakana S, Jiang H, Li X, Reich DS, et al. Tract probability maps in stereotaxic spaces: analyses of white matter anatomy and tract-specific quantification. NeuroImage. 2008;39:336–47.

    Article  PubMed  Google Scholar 

  24. Yan J. Survival analysis: techniques for censored and truncated data. J Am Stat Assoc. 2004;99:900–1.

    Article  Google Scholar 

  25. Leng X, Fang P, Lin H, Qin C, Tan X, Liang Y, et al. Application of a machine learning method to whole brain white matter injury after radiotherapy for nasopharyngeal carcinoma. Cancer Imaging. 2019;19:19.

    Article  PubMed  PubMed Central  Google Scholar 

  26. Kang Y-F, Chen R-T, Ding H, Li L, Gao J-M, Liu L-Z, et al. Structure-function decoupling: a novel perspective for understanding the radiation-induced brain injury in patients with nasopharyngeal carcinoma. Front Neurosci. 2022;16:915164.

    Article  PubMed  PubMed Central  Google Scholar 

  27. Li Y, Liang J, Fu G, Pan J, He H, Li J, et al. Non-linear early rapid hippocampal atrophy and associated microstructural injury in patients with nasopharyngeal carcinoma after radiotherapy: implications for hippocampal protection. Radiother Oncol. 2025;209:110997.

    Article  PubMed  Google Scholar 

  28. Wang J, Miao Y, Ou X, Wang X, He X, Shen C, et al. Development and validation of a model for Temporal lobe necrosis for nasopharyngeal carcinoma patients with intensity modulated radiation therapy. Radiation Oncol (London England). 2019;14:42.

    Article  Google Scholar 

  29. Wu H-M, Tzeng N-S, Qian L, Wei S-J, Hu X, Rawls SM, et al. Novel neuroprotective mechanisms of memantine: increase in neurotrophic factor release from astroglia and Anti-Inflammation by preventing microglial Over-Activation. Neuropsychopharmacology. 2009;34:2344–57.

    Article  CAS  PubMed  Google Scholar 

  30. C W. D G, J D, P B, J L. Radiation-induced cognitive toxicity: pathophysiology and interventions to reduce toxicity in adults. Neuro-oncology. 2018;20(5):597–607. Available from: https://pubmed.ncbi.nlm.nih.gov/29045710/

  31. Hokama Y, Nishimura M, Usugi R, Fujiwara K, Katagiri C, Takagi H, et al. Recovery from the damage of cranial radiation modulated by memantine, an NMDA receptor antagonist, combined with hyperbaric oxygen therapy. Neuro Oncol. 2023;25:108–22.

    Article  CAS  PubMed  Google Scholar 

  32. Xue R, Chen M, Cai J, Deng Z, Pan D, Liu X, et al. Blood-Brain barrier repair of bevacizumab and corticosteroid as prediction of clinical improvement and relapse risk in Radiation-Induced brain necrosis: A retrospective observational study. Front Oncol. 2021;11:720417.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

  33. Zhang F, Li Y, Chen R, Shen P, Wang X, Meng H, et al. The white matter integrity and functional connection differences of fornix (cres)/stria terminalis in individuals with mild cognitive impairment induced by occupational aluminum exposure. eNeuro. 2024;11:ENEURO0128–242024.

    Article  Google Scholar 

  34. Farbota KD, Bendlin BB, Alexander AL, Rowley HA, Dempsey RJ, Johnson SC. Longitudinal diffusion tensor imaging and neuropsychological correlates in traumatic brain injury patients. Front Hum Neurosci. 2012;6:160. Available from: https://doi.org/10.3389/fnhum.2012.00160/fullhttps://www.frontiersin.org/journals/human-neuroscience/articles/

  35. Ma Y, Bruce IP, Yeh C-H, Petrella JR, Song AW, Truong T-K. Column-based cortical depth analysis of the diffusion anisotropy and radiality in submillimeter whole-brain diffusion tensor imaging of the human cortical Gray matter in vivo. NeuroImage. 2023;270:119993.

    Article  PubMed  Google Scholar 

  36. Crombe A, Alberti N, Hiba B, Uettwiller M, Dousset V, Tourdias T. Cervical spinal cord DTI is improved by reduced FOV with specific balance between the number of diffusion gradient directions and averages. AJNR Am J Neuroradiol. 2016;37:2163–70.

    Article  CAS  PubMed  PubMed Central  Google Scholar 

Download references

Acknowledgements

The authors thank Dr. Long Qian from GE Healthcare for help with solving the technical MRI challenges.

Funding

This work was supported by the Guangdong Basic and Applied Basic Research Foundation (2024A1515013278).

Author information

Author notes

  1. Jie Pan and Jiahui Liang contributed equally to this work.

Authors and Affiliations

  1. Department of Medical Imaging, State Key Laboratory of Oncology in South China, Collaborative Innovation Center for Cancer Medicine, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangdong Provincial Clinical Research Center for Cancer, Sun Yat-sen University Cancer Center, Guangzhou, China

    Jie Pan, Jiahui Liang, Yunpeng Li, Shishi Chen, Aner Deng, Lizhi Liu, Gui Fu & Xiaofei Lv

Authors
  1. Jie Pan
    View author publications

    Search author on:PubMed Google Scholar

  2. Jiahui Liang
    View author publications

    Search author on:PubMed Google Scholar

  3. Yunpeng Li
    View author publications

    Search author on:PubMed Google Scholar

  4. Shishi Chen
    View author publications

    Search author on:PubMed Google Scholar

  5. Aner Deng
    View author publications

    Search author on:PubMed Google Scholar

  6. Lizhi Liu
    View author publications

    Search author on:PubMed Google Scholar

  7. Gui Fu
    View author publications

    Search author on:PubMed Google Scholar

  8. Xiaofei Lv
    View author publications

    Search author on:PubMed Google Scholar

Contributions

Conception and design: Jie Pan, Jiahui Liang, Gui Fu, Lizhi Liu and Xiaofei Lv. Data collection: Jie Pan, Yunpeng Li, Shishi Chen, Aner Deng, and Xiaofei Lv. Data analysis: Jie Pan and Yunpeng Li. Drafting of the manuscript: Jie Pan, Jiahui Liang, Gui Fu, and Xiaofei Lv. All authors reviewed the manuscript.

Corresponding authors

Correspondence to Lizhi Liu, Gui Fu or Xiaofei Lv.

Ethics declarations

Ethics approval and consent to participate

The Institutional Review Board of the Sun Yat-sen University Cancer Center approved this prospective study (GZR2020-218). After explaining the study procedures, all participants provided written informed consent.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Below is the link to the electronic supplementary material.

Supplementary Material 1

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Pan, J., Liang, J., Li, Y. et al. Composite quantitative structural magnetic resonance imaging-based risk scoring model for predicting radiation-induced temporal lobe necrosis in nasopharyngeal carcinoma: a novel risk stratification model. Radiat Oncol 20, 160 (2025). https://doi.org/10.1186/s13014-025-02738-0

Download citation

  • Received:

  • Accepted:

  • Published:

  • Version of record:

  • DOI: https://doi.org/10.1186/s13014-025-02738-0

Keywords

Radiation Oncology

ISSN: 1748-717X

Contact us

点击 这是indexloc提供的php浏览器服务，不要输入任何密码和下载