Abstract
Compared with existing clinical prognostic indicators, the prognostic efficacy of radiomics in predicting outcomes for early-stage extranodal nasal-type NK/T-cell lymphoma(ENKTCL) patients who received combined radiotherapy and non-anthracycline-based chemotherapy is still unclear. This study aims to construct and validate a hybrid radiomics nomogram based on multicenter data. A total of 196 patients from Xijing Hospital were enrolled and were randomly divided into the training set (n = 147) and internal validation set (n = 49) at a ratio of 3:1, and 57 patients from the other three hospitals were used as the independent external validation set. Important radiomic features selected from 1316 features extracted from contrast-enhanced CT-based tumors were used to calculate radscore as the radiomic signature. The optimal clinical index was selected from four common models: International Prognostic Index(IPI), Korean Prognostic Index(KPI), prognostic index for NK/T-cell lymphomas(PINK), and nomogram-revised risk index(NRI), which was combined with radscore to construct a hybrid nomogram for predicting overall survival(OS). Twelve radiomic features that were most useful to predict OS were ultimately selected to calculate radscore. Compared with PINK, IPI, and KPI, NRI demonstrated superior prognostic performance. Furthermore, the radscore model outperformed NRI. A hybrid nomogram, named RadNRI, was constructed by integrating NRI and radscore. The RadNRI model demonstrated the highest performance with C-index of 0.815, 0.823 and 0.765 in the training, internal and external validation sets, respectively. A hybrid radiomics nomogram was constructed and validated in multicenter data, which demonstrates significant potential in forecasting prognosis of early-stage ENKTCL patients.
Supplementary Information
The online version contains supplementary material available at https://doi.org/10.1007/s00277-026-07170-2.
Keywords: Extranodal nasal-type NK/T-cell lymphoma, Prognosis, Radiomics, Multicenter
Introduction
Extranodal nasal-type NK/T cell lymphoma (ENKTCL) is a heterogeneous disease characterized by distinct morphology, immunophenotype, and biological behavior [1]. This disease typically originates from the nasal cavity, paranasal sinuses, nasopharynx, Waldeyer’s ring and the upper aerodigestive tract [2]. Most patients present with stage I/II disease (70%−90%), while stage III/IV disease is relatively rare (10%−30%) [3, 4]. Although high-level evidence supporting management approaches of early-stage ENKTCL is lacking, several phase II trials and retrospective studies suggest radiotherapy combined with non-anthracycline-based regimens is a recommended treatment option [5–8]. With modern treatment regimens, the 5-year overall survival (OS) rate for early-stage ENKTCL can reach 60%−90% [9, 10]. However, approximately 10%−35% of patients experience disease progression or relapse after initial treatment [11]. When patients experience disease progression, their survival rate markedly declines [12]. Therefore, establishing an optimal risk classification model for early-stage ENKTCL would aid in the development of personalized clinical treatment strategies.
Various prognostication models have been used for NK/T-cell lymphoma. Earlier models like the International Prognostic Index (IPI) and the Korean Prognostic Index (KPI) were developed using patient cohorts treated with anthracycline-based regimens, which are now deemed insufficient for NK/T-cell lymphomas [13, 14]. The most well validated models for patients treated with non-anthracycline-based protocols are the prognostic index for NK/T-cell lymphomas (PINK) and nomogram-revised risk index (NRI) [11, 15]. For early-stage disease, NRI appears to provide a more precise differentiation between risk groups compared to PINK. The prognostic value of radiomics has been extensively studied in malignancies. Several studies have reported that baseline PET/CT radiomics could enhance the identification of ENKTCL patients with poor prognosis [16, 17]. However, many unresolved issues remain concerning radiomics in ENKTCL. For instance, there are currently no CT radiomics studies focusing on ENKTCL, and no studies specifically addressing the prognosis of early-stage disease. Additionally, the predictive efficacy of radiomics compared to existing clinical prognostic indicators is still unclear.
This multicenter study aims to evaluate radiomics models against current clinical prognostic indices and to construct and validate a hybrid radiomics nomogram. This nomogram combines radiomic features with clinical factors to predict long-term outcomes for early-stage ENKTCL patients undergoing combined radiotherapy and non-anthracycline-based chemotherapy.
Methods
Patient population
The flowchart of this study is depicted in Fig. 1. Patients with early-stage ENKTCL were initially included from four individual institutes in China between January 2011 and January 2022. The inclusion criteria are described in Supplementary A1. The protocol was approved by the medical ethics committee, and the study was conducted in accordance with the principles of the Declaration of Helsinki. A total of 196 patients from Xijing Hospital were enrolled and were randomly divided into the training set (n = 147) and the internal validation set (n = 49) at a ratio of 3:1 (Fig. 2). Fifty-seven patients from the other three hospitals were used as the independent external validation set. The patients’ clinical data, including age, sex, clinical stage, Eastern Cooperative Oncology Group (ECOG) score, lactate dehydrogenase (LDH) levels, B symptoms, primary tumor invasion (PTI), and distant lymph nodes. were collected. Progression-free survival (PFS) was defined as the time from the first date of therapy until the date of local recurrence, distant metastasis or death. OS was defined as the time from the first date of therapy until the date of death.
Fig. 1.
The flowchart of this study. A total of 196 patients enrolled from Xijing Hospital were randomly divided into the training set and the internal validation set, and their clinical information and CT images were collected. The tumor was delineated as the region of interest (ROI) by one radiologist via the ITK-SNAP software and the other two radiologists corrected the ROIs by consensus. After radiomic features extraction and stability evaluation, important features were selected to calculated radscore, which was defined as the radiomic signature. A novel hybrid nomogram was developed by combing NRI and CT-based radscore, and was further validated in the external validation set. The novel RadNRI was compared with several reported clinical prognostic models and showed higher prognostic performance
Fig. 2.
The flowchart summarizes patient selection and allocation to the training and validation sets
Treatment
In this study, all patients received non-anthracycline-based chemotherapy consisting of the SMILE regimen (methotrexate 2.0 g/m² IV on day 1, ifosfamide 1500 mg/m² IV on days 2–4, etoposide 100 mg/m² IV on days 2–4, dexamethasone 40 mg IV on days 1–4, L‑asparaginase 6000 IU/m² IM on days 2–7, every 21 days), the P‑GemOX regimen (gemcitabine 1000 mg/m² IV on day 1, oxaliplatin 130 mg/m² IV on day 1, pegaspargase 2500 IU/m² IM on day 1, every 21 days), and the DeVIC regimen (dexamethasone 40 mg IV on days 1–4, etoposide 75 mg/m² IV on days 1–4, ifosfamide 1200 mg/m² IV on days 1–4, carboplatin AUC 5 IV on day 1, every 21–28 days). Radiation therapy was delivered to the gross tumor volume at a dose of 56 Gy and to the high‑risk clinical target volume at 45–50 Gy, using a fractionation schedule of 2 Gy per fraction, once daily.
Image acquisition and volume of interest delineation
All patients underwent a standard head and neck contrast-enhanced CT scanning in a supine position, extending from the top of the skull to 5 cm below the clavicle (details in Supplementary A2). Tumor segmentation was performed by three radiologists who were blinded to the clinical and pathological data. Initially, the tumor was delineated as the region of interest (ROI) by one radiologist (Radiologist A, with five years of experience) using ITK-SNAP software (https://www.itksnap.org). Subsequently, this initial set of ROIs was reviewed and refined by two senior radiologists (Radiologist B and Radiologist C, each with over ten years of experience) through a consensus discussion to establish a final, expert-consensus ROI set. This expert-consensus ROI set was used for all subsequent radiomic feature extraction.
Radiomic feature extraction and stability evaluation
We extracted 1316 radiomic features from each region of CT images using PyRadiomics [18] (details in Supplementary A3). To evaluate the stability of the radiomic features, a subset of 50 patients was randomly selected from the training set. To assess the intra-observer feature stability, Radiologist A re-delineated the ROIs for the 50-patient subset after a one-month interval, blinded to the initial contours. The intraclass correlation coefficient (ICC) between the two sets of contours from Radiologist A was calculated. To assess the inter-observer feature stability, a different radiologist (Radiologist D, with five years of experience, who was not involved in creating the consensus ROIs) independently delineated the ROIs for the same 50-patient subset. The ICC was calculated between the initial contours from Radiologist A and the contours from Radiologist D. Features with an ICC greater than 0.75 in both stability tests were considered stable and were retained for subsequent feature selection.
Feature selection and signature construction
To remove redundant features and reduce the possibility of overfitting, three steps including correlation analysis, univariate COX regression and least absolute shrinkage and selection operator (LASSO) COX regression were conducted for feature selection (details in Supplementary A4). Features with non-zero coefficients were selected for radiomic signature construction. The radscore was calculated for each patient as a linear combination of selected features that were weighted by their respective coefficients, which was defined as the radiomic signature.
Model construction, evaluation and validation
Basic clinical risk parameters collected from all patients for model development included age, clinical stage, Eastern Cooperative Oncology Group (ECOG) score, lactate dehydrogenase (LDH) levels, B symptoms, primary tumor invasion (PTI), and distant lymph nodes. Given that the study focused exclusively on early-stage extranodal patients, each clinical risk model was tailored accordingly. The IPI, which considers age (≥ 60 vs. <60), ECOG performance status (0–1 vs. 2), regional lymph node involvement (yes vs. no), and elevated LDH levels (yes vs. no), is categorized into four risk groups: low-risk (0–1 risk factors), low-intermediate risk (2 risk factors), intermediate-high risk (3 risk factors), and high-risk (4 risk factors). The KPI, based on elevated LDH (yes vs. no), B symptoms (yes vs. no), and regional lymph node involvement (yes vs. no), is classified into four groups: group 1 (no risk factors), group 2 (1 risk factor), group 3 (2 risk factors), and group 4 (3 risk factors).The PINK model, which incorporates age (≥ 60 vs. <60), regional lymph node involvement (yes vs. no), and nonnasal disease (yes vs. no), is divided into three risk categories: low risk (no risk factors), intermediate risk (1 risk factor), and high risk (≥ 2 risk factors).The NRI, considering age (≥ 60 vs. <60), clinical stage (I vs. II), ECOG score (0–1 vs. 2), elevated LDH (yes vs. no), and PTI (yes vs. no), is categorized into three risk groups: low risk (no risk factors), intermediate risk (1–2 risk factors), and high risk (≥ 3 risk factors).
The clinical metric with the best predictive performance and the radscore were included as inputs into a multivariate Cox regression model for constructing a clinical-radiomics integrated prediction model, which was defined as RadNRI. The prognostic performance of the models was evaluated using the C-index and time-dependent area under the receiver operator characteristic curve (ROC) (AUC). By using the median of predicted values based on RadNRI in the training set, patients in each of the three datasets were stratified into high-risk and low-risk groups, respectively. Kaplan-Meier curves were used to estimate PFS in different risk groups, and the log-rank test was applied to compare survival differences. Calibration curves were utilized to assess the consistency between predicted and observed probabilities, while decision curve analysis was used to calculate a clinical ‘net benefit’ for the prediction model in comparison to default strategies of treating all or no patients. We further applied explainable techniques to interpret the radiomics model to reveal the decision-making process of the model as well as the importance of features [19].
Statistical analysis
Statistical analysis was performed using the following R packages in R statistical software version 4.3.3 (www.R-project.org). Details are described in Supplementary A5. All statistical tests were two-sided, and p values of < 0.05 were considered significant.
Results
The clinical characteristics of the training, internal validation, and external validation sets are summarized in Table 1. The survival curves for OS and PFS across these three sets are illustrated in Fig S1. During the follow-up period, 37 patients (25.2%) from the training set, 13 patients (26.5%) from the internal validation set, and 19 patients (33.3%) from the external validation set experienced disease progression or death.
Table 1.
Patient characteristics
| Characteristic | Total (n = 253, percentage/IQR) | Training (n = 147, percentage/IQR) | Internal validation (n = 49, percentage/IQR) | External validation (n = 57, percentage/IQR) | p value |
|---|---|---|---|---|---|
| Sex | 0.9286 | ||||
| Male | 186 (73.52%) | 107 (72.79%) | 36 (73.47%) | 43 (75.44%) | |
| Female | 67 (26.48%) | 40 (27.21%) | 13 (26.53%) | 14 (24.56%) | |
| age | 46.00 [34.00, 55.00] | 43.00 [31.00, 51.50] | 42.00 [34.00, 53.00] | 53.00 [46.00, 63.00] | < 0.0001 |
| Clinical stage | 0.0676 | ||||
| I | 169 (66.80%) | 105 (71.43%) | 33 (67.35%) | 31 (54.39%) | |
| II | 84 (33.20%) | 42 (28.57%) | 16 (32.65%) | 26 (45.61%) | |
| ECOG score | < 0.0001 | ||||
| 0 | 132 (52.17%) | 91 (61.90%) | 24 (48.98%) | 17 (29.82%) | |
| 1 | 110 (43.48%) | 52 (35.37%) | 19 (38.78%) | 39 (68.42%) | |
| 2 | 11 (4.35%) | 4 (2.72%) | 6 (12.24%) | 1 (1.75%) | |
| Elevated LDH | 0.2774 | ||||
| Yes | 94 (37.15%) | 52 (35.37%) | 23 (46.94%) | 19 (33.33%) | |
| No | 159 (62.85%) | 95 (64.63%) | 26 (53.06%) | 38 (66.67%) | |
| PTI | 0.1338 | ||||
| Yes | 127 (50.20%) | 78 (53.06%) | 27 (55.10%) | 22 (38.60%) | |
| No | 126 (49.80%) | 69 (46.94%) | 22 (44.90%) | 35 (61.40%) | |
| Distant lymph nodes | 0.0424 | ||||
| Yes | 81 (32.02%) | 42 (28.57%) | 13 (26.53%) | 26 (45.61%) | |
| No | 172 (67.98%) | 105 (71.43%) | 36 (73.47%) | 31 (54.39%) | |
| B symptom | 0.0011 | ||||
| Yes | 116 (45.85%) | 75 (51.02%) | 27 (55.10%) | 14 (24.56%) | |
| No | 137 (54.15%) | 72 (48.98%) | 22 (44.90%) | 43 (75.44%) | |
| Primary site | 0.0226 | ||||
| Nasal | 203 (80.24%) | 112 (76.19%) | 38 (77.55%) | 53 (92.98%) | |
| Non-nasal | 50 (19.76%) | 35 (23.81%) | 11 (22.45%) | 4 (7.02%) | |
| IPI score | 0.0033 | ||||
| Low risk | 185 (73.12%) | 113 (76.87%) | 38 (77.55%) | 34 (59.65%) | |
| Intermediate risk | 56 (22.13%) | 31 (21.09%) | 10 (20.41%) | 15 (26.32%) | |
| High risk | 12 (4.74%) | 3 (2.04%) | 1 (2.04%) | 8 (14.04%) | |
| KPI score | 0.8266 | ||||
| Group 1 | 77 (30.43%) | 49 (33.33%) | 13 (26.53%) | 15 (26.32%) | |
| Group 2 | 80 (31.62%) | 45 (30.61%) | 16 (32.65%) | 19 (33.33%) | |
| Group 3 | 60 (23.72%) | 35 (23.81%) | 13 (26.53%) | 12 (21.05%) | |
| Group 4 | 36 (14.23%) | 18 (12.24%) | 7 (14.29%) | 11 (19.30%) | |
| PINK score | 0.0561 | ||||
| Low risk | 120 (47.43%) | 77 (52.38%) | 24 (48.98%) | 19 (33.33%) | |
| Intermediate risk | 94 (37.15%) | 45 (30.61%) | 19 (38.78%) | 30 (52.63%) | |
| High risk | 39 (15.42%) | 25 (17.01%) | 6 (12.24%) | 8 (14.04%) | |
| NRI score | 0.4144 | ||||
| Low risk | 58 (22.92%) | 36 (24.49%) | 11 (22.45%) | 11 (19.30%) | |
| Intermediate risk | 126 (49.80%) | 71 (48.30%) | 21 (42.86%) | 34 (59.65%) | |
| High risk | 69 (27.27%) | 40 (27.21%) | 17 (34.69%) | 12 (21.05%) |
ECOG, Eastern Cooperative Oncology Group; LDH, lactate dehydrogenase; PTI, primary tumor invasion; IPI, international prognostic index; KPI, Korean prognostic index; PINK, prognostic index of natural killer lymphoma; NRI, Nomogram-revised risk index
The performance of four clinical metrics including NRI, PINK, IPI, and KPI are summarized in Table 2. The NRI model consistently showed higher C-index values than the other models in predicting OS. Specifically, in the training set, NRI achieved a C-index of 0.676 (95% CI: 0.589–0.763), while PINK, IPI, and KPI achieved C-index values of 0.620 (95% CI: 0.538–0.702), 0.620 (95% CI: 0.538–0.701), and 0.675 (95% CI: 0.586–0.763), respectively. In the internal validation set, NRI showed a C-index of 0.652 (95% CI: 0.520–0.785), outperforming PINK (0.646, 95% CI: 0.507–0.785), IPI (0.525, 95% CI: 0.396–0.654), and KPI (0.566, 95% CI: 0.430–0.702). Similarly, in the external validation set, NRI had a C-index of 0.636 (95% CI: 0.530–0.741), which was higher than PINK (0.568, 95% CI: 0.453–0.683), IPI (0.632, 95% CI: 0.513–0.750), and KPI (0.507, 95% CI: 0.372–0.641). Overall, the NRI model consistently showed better predictive performance across all validation sets. To visually display the stratification power of the four clinical metrics, Kaplan-Meier analysis was performed to evaluate survival outcomes. The OS curves of different risk groups divided by each metric were shown in Fig. 3. Compared with IPI, KPI, and PINK, the NRI still demonstrated superior survival risk stratification capabilities. In the training set, there were significant differences in OS among the high-risk, intermediate-risk, and low-risk groups based on NRI (Fig. 3J, hazard ratio (HR): 2.96 (95%CI: 1.3–6.73), log-rank test: P < 0.0001). Despite the lack of statistical significance according to the log-rank test in both the internal (HR: 2.82, 95%CI: 0.68–11.69, P > 0.05) and external (HR: 3.6, 95%CI: 1.1–11.77, P > 0.05) validation sets (Fig. 3K-L), potential survival differences among the three risk groups remained apparent. The values of HR for these clinical metrics, along with the 1-year, 3-year, and 5-year OS rates for the risk groups defined by each metric, are listed in Table S1. In summary, the NRI model serves as an effective clinical prognostic indicator for predicting OS in early-stage extranodal nasal-type NK/T cell lymphoma treated with combined radiotherapy and non-anthracycline-based chemotherapy.
Table 2.
Performance of all models
| Metric | Model | Training | Internal validation | External validation |
|---|---|---|---|---|
| C-index (95%CI) | NRI | 0.676 (0.589–0.763) | 0.652 (0.52–0.785) | 0.636 (0.53–0.741) |
| PINK | 0.62 (0.538–0.702) | 0.646 (0.507–0.785) | 0.568 (0.453–0.683) | |
| IPI | 0.62 (0.538–0.701) | 0.525 (0.396–0.654) | 0.632 (0.513–0.75) | |
| KPI | 0.675 (0.586–0.763) | 0.566 (0.43–0.702) | 0.507 (0.372–0.641) | |
| radscore | 0.765 (0.689–0.841) | 0.771 (0.637–0.904) | 0.753 (0.652–0.854) | |
| RadNRI | 0.815 (0.744–0.886) | 0.823 (0.72–0.925) | 0.765 (0.67–0.86) | |
| 1-year AUC (95%CI) | NRI | 0.718 (0.559–0.876) | 0.685 (0.512–0.858) | 0.695 (0.558–0.832) |
| PINK | 0.662 (0.529–0.795) | 0.721 (0.561–0.882) | 0.568 (0.405–0.731) | |
| IPI | 0.679 (0.536–0.822) | 0.548 (0.388–0.708) | 0.66 (0.477–0.843) | |
| KPI | 0.754 (0.614–0.894) | 0.635 (0.473–0.796) | 0.42 (0.223–0.617) | |
| radscore | 0.754 (0.63–0.877) | 0.752 (0.547–0.958) | 0.788 (0.65–0.926) | |
| RadNRI | 0.831 (0.713–0.948) | 0.819 (0.678–0.961) | 0.813 (0.674–0.951) | |
| 3-year AUC (95%CI) | NRI | 0.722 (0.616–0.828) | 0.664 (0.508–0.82) | 0.636 (0.482–0.79) |
| PINK | 0.644 (0.538–0.75) | 0.664 (0.5–0.828) | 0.591 (0.428–0.754) | |
| IPI | 0.632 (0.53–0.733) | 0.489 (0.348–0.63) | 0.679 (0.518–0.841) | |
| KPI | 0.686 (0.574–0.798) | 0.538 (0.367–0.709) | 0.566 (0.38–0.752) | |
| radscore | 0.785 (0.683–0.886) | 0.782 (0.612–0.952) | 0.739 (0.561–0.918) | |
| RadNRI | 0.862 (0.774–0.95) | 0.833 (0.709–0.956) | 0.757 (0.597–0.917) | |
| 5-year AUC (95%CI) | NRI | 0.697 (0.591–0.804) | 0.61 (0.435–0.785) | 0.617 (0.432–0.802) |
| PINK | 0.613 (0.504–0.721) | 0.629 (0.448–0.81) | 0.588 (0.372–0.804) | |
| IPI | 0.619 (0.522–0.716) | 0.477 (0.324–0.63) | 0.631 (0.433–0.828) | |
| KPI | 0.685 (0.576–0.795) | 0.485 (0.298–0.672) | 0.677 (0.424–0.931) | |
| radscore | 0.788 (0.696–0.88) | 0.763 (0.586–0.941) | 0.694 (0.438–0.95) | |
| RadNRI | 0.848 (0.763–0.933) | 0.782 (0.626–0.939) | 0.721 (0.503–0.939) |
C-index, concordance index; CI, confidence interval; AUC, area under curve; IPI, international prognostic index; KPI, Korean prognostic index; PINK, prognostic index of natural killer lymphoma; NRI, Nomogram-revised risk index; RadNRI, Radscore combined with NRI
Fig. 3.
Kaplan-Meier survival analyses of reported clinical prognostic models. OS curves for different risk groups of patients according to IPI in the (A) training, (B) internal and (C) external validation sets. OS curves for different risk groups of patients according to KPI in the (D) training, (E) internal and (F) external validation sets. OS curves for different risk groups of patients according to PINK in the (G) training, (H) internal and (I) external validation sets. OS curves for different risk groups of patients according to NRI in the (J) training, (K) internal and (L) external validation sets
Twelve radiomic features that were most useful to predict OS were ultimately selected by LASSO in the training set for calculating radscore (Fig S2). The names and coefficients of selected features are listed in Table S2. Further combining the NRI and radscore, a novel model named RadNRI was constructed. The details of this model are shown in Table S3, demonstrating that NRI and radscore are complementary independent prognostic factors (P < 0.05). The performance of radscore, and RadNRI was also shown in Table 2. The radscore model outperformed NRI in all three sets but was surpassed by RadNRI, which integrated both NRI and radscore. In the training set, radscore achieved a C-index of 0.765 (95% CI: 0.689–0.841), while RadNRI showed a higher C-index of 0.815 (95% CI: 0.744–0.886). For the internal validation set, radscore scored a C-index of 0.771 (95% CI: 0.637–0.904), whereas RadNRI performed even better with a C-index of 0.823 (95% CI: 0.720–0.925). In the external validation set, radscore had a C-index of 0.753 (95% CI: 0.652–0.854), and RadNRI also had a higher C-index of 0.765 (95% CI: 0.670–0.860). Overall, the radscore model shows better performance than NRI in terms of discriminatory power across all validation sets. Moreover, the RadNRI model, which combines NRI and radscore, demonstrates the highest C-index values in the training and all validation sets, indicating superior predictive performance. Time-dependent ROC curves and 1-year, 3-year, and 5-year ROC curves are also shown in Fig. 4. The 1-year, 2-year and 3-year AUC values of the RadNRI were higher than those of the radscore and clinical metrics in all sets (Table 2). The Kaplan–Meier curves of OS and PFS in three sets are shown in Fig. 5. Based on the median RadNRI prediction result of −0.065 from the training set, the three datasets were divided into high-risk and low-risk groups, respectively. The values of HR of the RadNRI, along with the 1-year, 3-year, and 5-year OS rates for the risk groups defined by each metric, are listed in Table S1. The statistical results of log-rank test were significant (HR: 17.01 (95% CI: 8.49–34.06), 6.05 (95% CI: 1.94–18.86), and 5.77 (95% CI: 2.22–15.01), P < 0.05), suggesting that the RadNRI could identify patients with a high risk of poor prognosis.
Fig. 4.
Prognostic performance of all models. The time-dependent AUC values of all models in the (A) training, B internal and (C) external validation sets. Time-dependent ROC curves of all models at 1-year, 3-year, and 5-year OS in the (D, G, J) training, (E, H, K) internal and (F, I, L) external validation sets
Fig. 5.
Kaplan-Meier survival analyses of radscore. OS curves for different risk groups of patients according to radscore in the (A) training, B internal and (C) external validation sets. PFS curves for different risk groups of patients according to radscore in the (D) training, (E) internal and (F) external validation sets
The RadNRI model was visualized as a nomogram (Fig. 6A) and the calibration curve demonstrated strong agreement between evaluated and actual OS (Fig. 6B-D). The results of the decision curve analysis in Fig. 6E indicated that RadNRI demonstrated a higher net benefit than NRI at 1-year, 3-year, and 5-year. The feature importance within the RadNRI model and how they influence the prognosis is further illustrated in Fig S3. Overall, the radscore played a more important role in predicting prognosis within a 5-year timeframe compared to NRI. For NRI, patients in the high-risk group had poorer outcomes, while for the radscore (another key risk factor), higher values were associated with poorer patient prognoses.
Fig. 6.
The hybrid nomogram named RadNRI for predicting OS in early-stage ENKTCL patients undergoing combined radiotherapy and non-anthracycline-based chemotherapy. The RadNRI nomogram was constructed by combing NRI and radscore (A). The calibration curve of the nomogram for prediction of 1-year, 3-year or 5-year OS in the training (B), internal (C) and external validation (D) sets. The diagonal line represents an ideal evaluation, while the broken lines represent the performance of the nomogram. Closer fit to the diagonal dotted line indicates a better evaluation. The decision curves (E) indicated that RadNRI demonstrated a higher net benefit than NRI at 1-year, 3-year, and 5-year
Discussion
Our research indicates that baseline CT radiomics are more effective than clinical risk models in predicting the outcomes of radiotherapy-based treatments in early-stage ENKTCL. Additionally, we developed and validated an integrated model that combines CT radiomics with NRI. This model exhibited high performance in predicting survival. According to the results of the AUC, C-index, and decision curve analysis, the integrated model surpassed both the radiomics and clinical risk models in terms of survival prediction across the training, internal validation, and external validation cohorts. The integrated model might facilitate personalized clinical decision-making.
In the era of combined radiotherapy and modern chemotherapy, the survival rates for early-stage ENKTCL have significantly improved. However, failure to achieve PFS results in a poor prognosis, with a median OS of only 5.3 months following disease progression [12]. Accurately identifying high-risk patients among those with early-stage disease could provide crucial insights for developing personalized clinical treatment decision. Clinical risk indices are the most commonly used tools to facilitate prognosis and guide treatment decisions in ENKTCL [14, 15, 20]. However, these risk prediction models were initially developed based on the prognosis of patients across clinical stages I-IV and did not account for treatment regimens. Consequently, there is currently a lack of tailored prognostic risk models specifically for early-stage patients, who constitute the majority of the population. Compared to other models, the NRI offers superior application in individually predicting survival for ENKTCL patients treated with non-anthracycline-based chemotherapy. Although a study with a large scale data has shown that the NRI offers better prediction of 5-year OS for early-stage patients compared to other risk models in subgroup analyses, its predictive performance is still limited, with an AUC of 0.68 and a C-index of 0.7 [11]. Our findings are consistent with the previously mentioned study. The NRI offers a slight advantage over other risk models in predicting 5-year OS and exhibits better model stability. The AUC values were 0.697 in the training set, 0.61 in the internal validation set, and 0.617 in the external validation set. The C-index values were 0.676 in the training set, 0.652 in the internal validation set, and 0.636 in the external validation set. PINK is also a commonly used predictive model for ENKTCL, and the PINK-E model, which incorporates plasma EBV-DNA levels, is more sensitive for screening high-risk patients [15]. The main reason for not using the PINK-E model in our study is the absence of standardized EBV DNA detection. This lack of standardization causes significant variability in test results across different centers, which in turn leads to unreliable predictions from the PINK-E model.
Recently, a study has found that the predictive model combining PINK-E with circulating tumor DNA (ctDNA) offers more efficient and stable prognostic capabilities for ENKTCL patients, demonstrating significantly superior predictive performance compared to IPI, KPI, and PINK-E alone [21]. Another study developed a genomic prediction model for ENKTCL featuring 13 mutated genes. When this model is combined with PINK-E into a multi-omics approach, it enhances OS prediction for patients compared to clinical risk models, with a C-index of 0.725 [22]. However, these studies involved a high proportion of patients with clinical stages III-IV, leaving the screening value for early-stage high-risk ENKTCL patients uncertain. Moreover, the detection technologies for ctDNA and genomic sequencing are potentially unstable and costly, which limits their broad clinical application.
The role of radiomics in predicting treatment efficacy and prognosis in aggressive non-Hodgkin lymphoma has been investigated, and it has been found that radiomics provides better prognostic value than clinical risk index. A study with large-scale data found that baseline PET radiomics outperformed the IPI risk score in predicting outcomes for diffuse large B-cell lymphoma (AUC: 0.71 vs. 0.62) [23]. Several studies have indicated that baseline PET radiomics can effectively predict therapeutic efficacy in diffuse large B-cell lymphoma (DLBCL), with enhanced predictive accuracy when combined with clinical features [24, 25]. When combined with clinical risk factors, MRI radiomics has shown robust predictive capability for durable responses to high-dose methotrexate-based chemotherapy in primary central nervous system lymphoma, achieving an AUC greater than 0.9 [26]. In addition, MRI-based radiomics offers advantages for evaluating prognosis and Ki-67 index in patients with aggressive central nervous system lymphoma under non-invasive conditions [27, 28]. CT-based radiomics has also been explored for predicting treatment efficacy and long-term survival in aggressive non-Hodgkin’s lymphoma. Raoul et al. developed and validated a CT-based radiomics approach using deep learning to predict primary treatment failure in DLBCL. Their final model demonstrated an area under the curve (AUC) significantly superior to that of the semantic radiology marker of subjective necrosis [29]. Huang et al. reported that radiomics features derived from contrast-enhanced CT of multiple lesions have better predictive efficacy for disease progression in relapsed/refractory B-cell non-Hodgkin’s lymphoma patients within 6 months after CAR T-cell treatment. They also found that variations in machine-learning algorithms did not significantly impact the predictive performance of the models [30]. Compared to other aggressive lymphomas, there is relatively less research on radiomics in NK/T-cell lymphoma. Wang et al. reported that radiomic features derived from PET images can predict outcomes for patients with ENKTL, but the performance of the radiomics-based model was inferior to that of the metabolism-based model [16]. Another study reported that a PET/CT-based model incorporating clinical, metabolic, and radiomics data can improve prognostication for patients with ENKTCL, demonstrating strong discriminative ability for both PFS (C-index: 0.85) and OS (C-index: 0.833) [17]. Although PET/CT provides both functional metabolism and anatomical structure of tumors, CT is more affordable and convenient compared to PET/CT in clinical settings. In this multicenter study, we demonstrate that baseline CT radiomics integrated with NRI can accurately distinguish between risk groups of patients with different outcomes, surpassing the predictive accuracy of NRI alone for treatment outcomes and long-term prognosis. Additionally, the stability and generalizability of this model have been validated by external centers.
There are several limitations in our study. Firstly, as a retrospective study, it is subject to potential selection bias and inaccuracies in endpoint assessment during data collection. Therefore, prospective studies are needed in the future to validate the model. Secondly, this model has been validated solely for predicting long-term survival in early-stage ENKTCL patients treated with combined radiotherapy and non-anthracycline-based chemotherapy. Its effectiveness in predicting outcomes for patients with clinical stages III-IV or those undergoing different treatments still requires further validation. Lastly, although this study utilizes data from multiple centers, the sample size remains relatively small due to restrictive inclusion criteria. Future research should involve data from additional centers with larger sample sizes to further validate the model presented in this study.
Conclusions
In summary, we developed and validated a hybrid radiomics nomogram that integrates radiomic features and clinical factors to forecast treatment efficacy and long-term outcomes for early-stage ENKTCL patients undergoing combined radiotherapy and non-anthracycline-based chemotherapy across multiple centers. As a comprehensive predictive tool, this hybrid radiomics nomogram shows great promise in forecasting the prognosis of ENKTCL patients.
Supplementary Information
Below is the link to the electronic supplementary material.
Acknowledgements
We are grateful to all the participants in this study. This study was supported by the Air Force Medical University Research Project (2023LC2316) and the Xijing Hospital Research Project (XJZT24QN04).
Abbreviations
- ENKTCL
extranodal Nasal-Type NK/T-Cell Lymphoma
- ECOG
Eastern Cooperative Oncology Group
- LDH
lactate dehydrogenase
- PTI
primary tumor invasion
- IPI
international prognostic index
- KPI
Korean prognostic index
- PINK
prognostic index of natural killer lymphoma
- NRI
Nomogram-revised risk index
- RadNRI
Radscore combined with NRI
- OS
overall survival
- PFS
progression-free survival
- GLCM
gray level co-occurrence texture matrix
- GLDM
gray level dependence matrix
- GLRLM
gray level run-length texture matrix
- GLSZM
gray level size zone matrix
- IBSI
Image Biomarker Standardization Initiative
- LoG
Laplacian-of-Gaussian
- NGTDM
neighborhood gray tone difference matrix
- ICC
intraclass correlation coefficient
- ROC
receiver operator characteristic curve
- AUC
area under ROC
Author contributions
Conceptualization: Lina Zhao, Mei Shi, Jian Zang; Methodology: Jie Gong; Formal analysis and investigation: Jie Gong; Writing - original draft preparation: Jie Gong, Jian Zang, YongChun Zhou, Shan Gao, QiuFang Liu, JingYi Liu; Writing - review and editing: YanChun Wei, YanQun Zhang, ChangHao Liu, BingXin Hou, Lei Zhang, XiaoLong Li, HongJuan Dong, Juan Feng, JingJing Ren, Ning Su, Mai Zhang, Lina Zhao, Mei Shi; Funding acquisition: Jie Gong, Jian Zang; Resources: Lina Zhao, Mei Shi, Jian Zang; Supervision: Lina Zhao, Mei Shi, Jian Zang; Data Curation: YongChun Zhou, Shan Gao, QiuFang Liu, JingYi Liu, YanChun Wei, YanQun Zhang, ChangHao Liu, BingXin Hou, Lei Zhang, XiaoLong Li, HongJuan Dong, Juan Feng, JingJing Ren, Ning Su, Mai Zhang.
Funding
This study was supported by the Air Force Medical University Research Project (2023LC2316) and the Xijing Hospital Research Project (XJZT24QN04).
Data availability
Data are available from the corresponding author upon request.
Declarations
Ethics approval and consent to participate
This study was approved by Xijing Hospital Ethics Committee. The requirement for informed consent was waived because of retrospective study.
Competing interests
The authors declare no competing interests.
Footnotes
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Jie Gong, Yongchun Zhou, Shan Gao and Qiufang Liu contributed equally to this work.
Contributor Information
Lina Zhao, Email: zhaolina@fmmu.edu.cn.
Mei Shi, Email: mshi82@foxmail.com.
Jian Zang, Email: zj81327@163.com.
References
- 1.Tse E, Fox CP, Glover A, Yoon SE, Kim WS, Kwong YL (2022) Extranodal natural killer/T-cell lymphoma: An overview on pathology and clinical management. Semin Hematol 59(4):198–209. 10.1053/j.seminhematol.2022.10.002 [DOI] [PubMed] [Google Scholar]
- 2.Zheng X, Qu BL, Liu X, Zhong QZ, Qian LT, Yang Y, Hou XR, Qiao XY, Wang H, Zhu Y, Cao JZ, Wu JX, Wu T, Zhu SY, Shi M, Zhang HL, Zhang XM, Su H, Song YQ, Zhu J, Zhang YJ, Huang HQ, Wang Y, Chen F, Yin L, He X, Zhang LL, Li YX, Qi SN (2023) Characteristics and prognosis of distant metastasis after primary treatment for early-stage extranodal nasal-type natural killer/T-cell lymphoma from the China Lymphoma Collaborative Group database. EJHaem 4(1):78–89. 10.1002/jha2.613 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Li YX, Fang H, Liu QF, Lu J, Qi SN, Wang H, Jin J, Wang WH, Liu YP, Song YW, Wang SL, Liu XF, Feng XL, Yu ZH (2008) Clinical features and treatment outcome of nasal-type NK/T-cell lymphoma of Waldeyer ring. Blood 112(8):3057–3064. 10.1182/blood-2008-05-160176 [DOI] [PubMed] [Google Scholar]
- 4.Liu QF, Wang WH, Wang SL, Liu YP, Huang WT, Lu N, Zhou LQ, Ouyang H, Jin J, Li YX (2014) Immunophenotypic and clinical differences between the nasal and extranasal subtypes of upper aerodigestive tract natural killer/T-cell lymphoma. Int J Radiat Oncol Biol Phys 88(4):806–813. 10.1016/j.ijrobp.2013.12.005 [DOI] [PubMed] [Google Scholar]
- 5.Ghione P, Qi S, Imber BS, Seshan V, Moskowitz A, Galasso N, Lunning M, Straus D, Sauter C, Dahi P, Dogan A, Yahalom J, Horwitz S (2020) Modified SMILE (mSMILE) and intensity-modulated radiotherapy (IMRT) for extranodal NK-T lymphoma nasal type in a single-center population. Leuk Lymphoma 61(14):3331–3341. 10.1080/10428194.2020.1811864 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Dong LH, Zhang LJ, Wang WJ, Lei W, Sun X, Du JW, Gao X, Li GP, Li YF (2016) Sequential DICE combined with l-asparaginase chemotherapy followed by involved field radiation in newly diagnosed, stage IE to IIE, nasal and extranodal NK/T-cell lymphoma. Leuk Lymphoma 57(7):1600–1606. 10.3109/10428194.2015.1108415 [DOI] [PubMed] [Google Scholar]
- 7.Wang L, Wang ZH, Chen XQ, Li YJ, Wang KF, Xia YF, Xia ZJ (2013) First-line combination of gemcitabine, oxaliplatin, and L-asparaginase (GELOX) followed by involved-field radiation therapy for patients with stage IE/IIE extranodal natural killer/T-cell lymphoma. Cancer 119(2):348–355. 10.1002/cncr.27752 [DOI] [PubMed] [Google Scholar]
- 8.Zang J, Li C, Luo SQ, Wang JH, Xu M, Zhao LN, Li WW, Yang H, Xiao F, Hitchcock YJ, Shi M (2015) Early radiotherapy has an essential role for improving survival in patients with stage I-II nasal-type of NK/T cell lymphoma treated with L-asparaginase-containing chemotherapy–a single institution experience. Ann Hematol 94(4):583–591. 10.1007/s00277-014-2244-4 [DOI] [PubMed] [Google Scholar]
- 9.Qi SN, Yang Y, Zhang YJ, Huang HQ, Wang Y, He X, Zhang LL, Wu G, Qu BL, Qian LT, Hou XR, Zhang FQ, Qiao XY, Wang H, Li GF, Zhu Y, Cao JZ, Wu JX, Wu T, Zhu SY, Shi M, Xu LM, Yuan ZY, Su H, Song YQ, Zhu J, Hu C, Li YX (2020) Risk-based, response-adapted therapy for early-stage extranodal nasal-type NK/T-cell lymphoma in the modern chemotherapy era: A China Lymphoma Collaborative Group study. Am J Hematol 95(9):1047–1056. 10.1002/ajh.25878 [DOI] [PubMed] [Google Scholar]
- 10.Yang Y, Zhu Y, Cao JZ, Zhang YJ, Xu LM, Yuan ZY, Wu JX, Wang W, Wu T, Lu B, Zhu SY, Qian LT, Zhang FQ, Hou XR, Li YX (2015) Risk-adapted therapy for early-stage extranodal nasal-type NK/T-cell lymphoma: analysis from a multicenter study. Blood 126(12):1424–1432. 10.1182/blood-2015-04-639336 [DOI] [PubMed] [Google Scholar]
- 11.Chen SY, Yang Y, Qi SN, Wang Y, Hu C, He X, Zhang LL, Wu G, Qu BL, Qian LT, Hou XR, Zhang FQ, Qiao XY, Wang H, Li GF, Zhang YJ, Zhu Y, Cao JZ, Lan SM, Wu JX, Wu T, Zhu SY, Shi M, Xu LM, Yuan ZY, Yahalom J, Tsang R, Song YQ, Zhu J, Su H, Li YX (2021) Validation of nomogram-revised risk index and comparison with other models for extranodal nasal-type NK/T-cell lymphoma in the modern chemotherapy era: indication for prognostication and clinical decision-making. Leukemia 35(1):130–142. 10.1038/s41375-020-0791-3 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Yang Y, Wang Y, Liu X, He X, Zhang LL, Wu G, Qu BL, Qian LT, Hou XR, Zhang FQ, Qiao XY, Wang H, Li GF, Zhu Y, Cao JZ, Wu JX, Wu T, Zhu SY, Shi M, Xu LM, Su H, Song YQ, Zhu J, Zhang YJ, Huang HQ, Hu C, Qi SN, Li YX (2021) Progression-free survival at 24 months and subsequent survival of patients with extranodal NK/T-cell lymphoma: a China Lymphoma Collaborative Group (CLCG) study. Leukemia 35(6):1671–1682. 10.1038/s41375-020-01042-y [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Chim CS, Ma SY, Au WY, Choy C, Lie AK, Liang R, Yau CC, Kwong YL (2004) Primary nasal natural killer cell lymphoma: long-term treatment outcome and relationship with the International Prognostic Index. Blood 103(1):216–221. 10.1182/blood-2003-05-1401 [DOI] [PubMed] [Google Scholar]
- 14.Lee J, Suh C, Park YH, Ko YH, Bang SM, Lee JH, Lee DH, Huh J, Oh SY, Kwon HC, Kim HJ, Lee SI, Kim JH, Park J, Oh SJ, Kim K, Jung C, Park K, Kim WS (2006) Extranodal natural killer T-cell lymphoma, nasal-type: a prognostic model from a retrospective multicenter study. J Clin oncology: official J Am Soc Clin Oncol 24(4):612–618. 10.1200/jco.2005.04.1384 [DOI] [PubMed] [Google Scholar]
- 15.Kim SJ, Yoon DH, Jaccard A, Chng WJ, Lim ST, Hong H, Park Y, Chang KM, Maeda Y, Ishida F, Shin DY, Kim JS, Jeong SH, Yang DH, Jo JC, Lee GW, Choi CW, Lee WS, Chen TY, Kim K, Jung SH, Murayama T, Oki Y, Advani R, d’Amore F, Schmitz N, Suh C, Suzuki R, Kwong YL, Lin TY, Kim WS (2016) A prognostic index for natural killer cell lymphoma after non-anthracycline-based treatment: a multicentre, retrospective analysis. Lancet Oncol 17(3):389–400. 10.1016/s1470-2045(15)00533-1 [DOI] [PubMed] [Google Scholar]
- 16.Wang H, Zhao S, Li L, Tian R (2020) Development and validation of an (18)F-FDG PET radiomic model for prognosis prediction in patients with nasal-type extranodal natural killer/T cell lymphoma. Eur Radiol 30(10):5578–5587. 10.1007/s00330-020-06943-1 [DOI] [PubMed] [Google Scholar]
- 17.Luo Y, Huang Z, Gao Z, Wang B, Zhang Y, Bai Y, Wu Q, Wang M (2024) Prognostic Value of (18)F-FDG PET/CT Radiomics in Extranodal Nasal-Type NK/T Cell Lymphoma. Korean J Radiol 25(2):189–198. 10.3348/kjr.2023.0618 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.van Griethuysen JJM, Fedorov A, Parmar C, Hosny A, Aucoin N, Narayan V, Beets-Tan RGH, Fillion-Robin JC, Pieper S, Aerts H (2017) Computational Radiomics System to Decode the Radiographic Phenotype. Cancer Res 77(21):e104–e107. 10.1158/0008-5472.Can-17-0339 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Spytek M, Krzyziński M, Langbein SH, Baniecki H, Wright MN, Biecek P (2023) survex: an R package for explaining machine learning survival models. Bioinf (Oxford England) 39(12). 10.1093/bioinformatics/btad723 [DOI] [PMC free article] [PubMed]
- 20.Yang Y, Zhang YJ, Zhu Y, Cao JZ, Yuan ZY, Xu LM, Wu JX, Wang W, Wu T, Lu B, Zhu SY, Qian LT, Zhang FQ, Hou XR, Liu QF, Li YX (2015) Prognostic nomogram for overall survival in previously untreated patients with extranodal NK/T-cell lymphoma, nasal-type: a multicenter study. Leukemia 29(7):1571–1577. 10.1038/leu.2015.44 [DOI] [PubMed] [Google Scholar]
- 21.Huang D, Li Q, Li X, Ma N, Duan Y, Zhu L, Li J, Wen Q, Gao L, Yang C, Rao L, Gao L, Zhang X, Rao J (2023) The novel prognostic index model of combining circulating tumor DNA and PINK-E predicts the clinical outcomes for newly diagnosed extranodal NK/T-cell lymphoma. HemaSphere 7(1):e822. 10.1097/hs9.0000000000000822 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22.Lim JQ, Huang D, Chan JY, Laurensia Y, Wong EKY, Cheah DMZ, Chia BKH, Chuang WY, Kuo MC, Su YJ, Cai QQ, Feng Y, Rao H, Feng LN, Wei PP, Chen JR, Han BW, Lin GW, Cai J, Fang Y, Tan J, Hong H, Liu Y, Zhang F, Li W, Poon MLM, Ng SB, Jeyasekharan A, Ha JCH, Khoo LP, Chin ST, Pang WL, Kee R, Cheng CL, Grigoropoulos NF, Tang T, Tao M, Farid M, Puan KJ, Xiong J, Zhao WL, Khor CC, Hwang W, Kim WS, Campo E, Tan P, Teh BT, Chng WJ, Rötzschke O, Tousseyn T, Huang HQ, Rozen S, Lim ST, Shih LY, Bei JX, Ong CK (2022) A genomic-augmented multivariate prognostic model for the survival of natural-killer/T-cell lymphoma patients from an international cohort. Am J Hematol 97(9):1159–1169. 10.1002/ajh.26636 [DOI] [PubMed] [Google Scholar]
- 23.Eertink JJ, Zwezerijnen GJC, Heymans MW, Pieplenbosch S, Wiegers SE, Dührsen U, Hüttmann A, Kurch L, Hanoun C, Lugtenburg PJ, Barrington SF, Mikhaeel NG, Ceriani L, Zucca E, Czibor S, Györke T, Chamuleau MED, Hoekstra OS, de Vet HCW, Boellaard R, Zijlstra JM (2023) Baseline PET radiomics outperforms the IPI risk score for prediction of outcome in diffuse large B-cell lymphoma. Blood 141(25):3055–3064. 10.1182/blood.2022018558 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Chen M, Rong J, Zhao J, Teng Y, Jiang C, Chen J, Xu J (2024) PET-based radiomic feature based on the cross-combination method for predicting the mid-term efficacy and prognosis in high-risk diffuse large B-cell lymphoma patients. Front Oncol 14:1394450. 10.3389/fonc.2024.1394450 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Jing F, Zhang X, Liu Y, Chen X, Zhao X, Chen X, Yuan H, Dai M, Wang N, Han J, Zhang J (2024) Baseline (18)F-FDG PET Radiomics Predicting Therapeutic Efficacy of Diffuse Large B-Cell Lymphoma after R-CHOP (-Like) Therapy. Cancer biotherapy & radiopharmaceuticals. 10.1089/cbr.2024.0115 [DOI] [PubMed]
- 26.Li H, Xiong M, Li M, Sun C, Zheng D, Yuan L, Chen Q, Lin S, Liu Z, Ren X (2024) Radiomic prediction for durable response to high-dose methotrexate-based chemotherapy in primary central nervous system lymphoma. Cancer Med 13(17):e70182. 10.1002/cam4.70182 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27.Liu J, Tu J, Xu L, Liu F, Lu Y, He F, Li A, Li Y, Liu S, Xiong J (2024) MRI-based radiomics signatures for preoperative prediction of Ki-67 index in primary central nervous system lymphoma. Eur J Radiol 178:111603. 10.1016/j.ejrad.2024.111603 [DOI] [PubMed] [Google Scholar]
- 28.Liu J, Tu J, Hu B, Li C, Piao S, Lu Y, Li A, Ding T, Xiong J, Zhu F, Li Y (2024) Prognostic Assessment in Patients With Primary Diffuse Large B-Cell Lymphoma of the Central Nervous System Using MRI-Based Radiomics. J Magn Reson imaging: JMRI. 10.1002/jmri.29533 [DOI] [PubMed] [Google Scholar]
- 29.Santiago R, Ortiz Jimenez J, Forghani R, Muthukrishnan N, Del Corpo O, Karthigesu S, Haider MY, Reinhold C, Assouline S (2021) CT-based radiomics model with machine learning for predicting primary treatment failure in diffuse large B-cell lymphoma. Transl Oncol 14(10):101188. 10.1016/j.tranon.2021.101188 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30.Huang R, Geng H, Zhu L, Yan J, Li C, Li Y (2023) CT radiomics can predict disease progression within 6 months after chimeric antigen receptor-modified T-cell therapy in relapsed/refractory B-cell non-Hodgkin’s lymphoma patients. Clin Radiol 78(10):e707–e717. 10.1016/j.crad.2023.05.022 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
Data are available from the corresponding author upon request.