Primary central nervous system lymphoma (PCNSL) is an uncommon and aggressive malignancy that accounts for approximately 1-2% of all non-Hodgkin lymphoma (NHL) cases and about 3-4% of all CNS tumors. The overwhelmingly dominant histological type is diffuse large B-cell lymphoma (DLBCL), which is the same aggressive subtype seen in systemic lymphoma but confined to the brain, spinal cord, eyes, or cerebrospinal fluid without detectable systemic involvement. Because PCNSL lacks a universally accepted salvage treatment regimen, initial therapy selection and early identification of patients at risk for treatment failure are critical clinical priorities.
Treatment landscape: The backbone of PCNSL therapy is high-dose methotrexate (MTX), typically administered at 8 g/m2 for 4 to 6 induction cycles. Adjuvant whole-brain radiotherapy (WBRT) following chemotherapy improves survival versus chemotherapy alone. Despite these approaches, approximately 50% of PCNSL patients experience relapsed or refractory (R/R) disease, which carries a markedly worse prognosis. In this study's cohort, 20 of 46 patients (43.5%) developed R/R disease, and the mortality rate among R/R patients was 60%, compared to 30.8% in those without relapse (p = 0.047).
The limitation of conventional MRI assessment: Conventional MRI characteristics (tumor size, T2 hyperintensity, infratentorial location, apparent diffusion coefficient values) have been identified as prognostic factors in PCNSL, but these assessments are largely qualitative and subject to inter-rater variability. The lack of reproducibility in visual MRI interpretation makes it difficult to use these factors reliably in clinical decision-making. Radiomics offers a quantitative alternative by converting medical images into hundreds of numerical features that can be analyzed systematically.
This 2024 paper from Chi Mei Medical Center and E-Da Hospital in Taiwan specifically investigates whether pretreatment MRI-based radiomics combined with machine learning can predict R/R disease in PCNSL before treatment begins. If successful, such a model could guide clinicians in selecting more aggressive upfront treatment or consolidation strategies for high-risk patients.
Radiomic studies specifically targeting prognosis prediction in PCNSL are exceptionally rare, largely because the disease itself is uncommon. All published studies have been retrospective and included small datasets. A critical gap noted by the authors is that no prior work has directly compared MRI-based radiomic machine learning models with quantitative ADC measurements for predicting R/R in PCNSL, which is precisely what this study addresses.
Destito et al. (2023): Used an ensemble of ML models including Extra Tree Classifier, SVM, logistic regression, random forest, and k-nearest neighbors to predict overall survival (OS) and progression-free survival (PFS) in 23 PCNSL patients. Their radiomics-based model achieved AUC of 0.86 and 0.84 for OS and PFS prediction respectively, outperforming clinical factor-based models. However, the study only used T1WI and T2WI sequences for feature extraction, limiting coverage of the full MRI information available.
Nenning et al. (2023): Published the largest PCNSL radiomic study to date, with 191 patients across nine institutions in Austria. Using contrast-enhanced T1WI textural features combined with SVM, they reported that a radiomic risk score is a robust and complementary predictor of survival that correlates with DNA methylation patterns in PCNSL tumors. This study provides important evidence that radiomic features capture biologically meaningful tumor characteristics, not just imaging noise. However, it used only a single MRI sequence for feature extraction.
Ali et al. (2020) and She et al. (2023): Ali et al. used CE T1 and T2-FLAIR radiomic features in an SVM framework to predict MTX-based treatment response in 47 patients, achieving 81.1% accuracy and AUC of 0.81. She et al. pioneered deep learning for PCNSL prognosis prediction using a 3D-ResNet model on 56 patients, obtaining AUC of 0.81, accuracy of 87%, and precision of 88% on CE T1WI alone. The present study builds on these works by integrating three MRI sequences (CE T1WI, T2WI, and T2 FLAIR) and directly comparing multiple ML algorithms against each other and against ADC.
The study enrolled 46 consecutive PCNSL patients (20 male, 26 female, median age 64 years) from two institutions between October 2005 and November 2020. Inclusion required postoperative histopathological confirmation of PCNSL (all cases were DLBCL), complete pretreatment and post-treatment brain MRI follow-up, absence of systemic involvement confirmed by bone marrow biopsy and PET/CT scans, and no immunocompromised conditions or HIV infection. All patients received consistent MTX-based induction chemotherapy, making the cohort relatively homogeneous in terms of treatment. Additionally, 18 patients also received rituximab as part of induction, and 20 patients received WBRT (with 2 receiving both WBRT and autologous stem cell transplantation due to R/R disease).
MRI acquisition: Preoperative brain MRI was performed using 1.5-T scanners in 42 cases (Siemens Avanto, Siemens Aera, or GE Signa) and a 3-T scanner (GE Discovery MR750) in 4 cases, all with eight-channel head coils. The imaging protocol included spin-echo T1-weighted imaging (T1WI), fast spin-echo T2-weighted imaging (T2WI), T2-FLAIR, diffusion-weighted imaging (DWI), ADC mapping, and contrast-enhanced T1WI (CE-T1WI) using either gadobutrol or gadoterate meglumine at 0.1 mmol/kg. DWI was acquired at b-values of 1000 and 1500 s/mm2 along x, y, and z axes.
Tumor segmentation: Rather than using deep learning segmentation (which requires large annotated datasets the authors did not have), the team opted for a semi-automatic fuzzy C-means (FCM) clustering algorithm applied to CE-T1WI. An experienced neuroradiologist (12 years of experience) determined the tumor location and slice range, and the algorithm automatically segmented the lesion ROI on each slice, combining them into a 3D volume of interest (VOI) using connected-component labeling. The VOI was then co-registered to T2WI and FLAIR images using ITK-SNAP 4.0 for consistent cross-sequence feature extraction. This approach prioritizes reproducibility given the small dataset size.
Feature extraction: Using PyRadiomics 3.0 in Python, 107 radiomic features per MRI sequence were extracted from the tumor VOI: 14 shape-based features characterizing lesion geometry, 18 first-order statistical features describing voxel intensity distributions, and 75 texture features quantifying intratumoral heterogeneity. All features were Z-score normalized (subtracting the mean, dividing by the standard deviation) to ensure comparability across patients and sequences.
With 107 features extracted from each of three MRI sequences, the total feature space was large relative to the 46-patient dataset. To avoid overfitting, the authors used SVM with a Gaussian (radial basis function) kernel and a sequential forward selection process to identify the five most informative features across sequences. This selection used 10-fold cross-validation with 500 Monte Carlo repetitions to reliably estimate generalization error and ensure the identified features are robustly predictive rather than coincidentally associated. The five selected features were: GLSZM large area low gray level emphasis, first-order kurtosis, original first-order minimum, original GLCM cluster tendency, and original GLRLM gray level non-uniformity normalized. These features collectively capture lesion-level texture heterogeneity, intensity distribution extremes, and spatial co-occurrence patterns within the tumor.
Five ML algorithms: Using these five selected features, five predictive models were built and evaluated with 10-fold cross-validation: support vector machine (SVM) with a Gaussian kernel, k-nearest neighbors (KNN), linear discriminant analysis (LDA), naive Bayes (NB), and decision trees (DT). All models were implemented in MATLAB's Classification Learner app using default settings to maximize reproducibility and minimize risk of overfitting through hyperparameter tuning on a small dataset. A radiomics score was calculated for each patient using a 0.5 threshold to classify R/R versus non-R/R status.
ADC measurement as a comparator: To contextualize the ML model performance, ADC values were independently measured from DWI (b = 1000 s/mm2) by two blinded neuroradiologists (12 and 16 years of experience, respectively). A circular ROI with area ranging from 22 to 64 mm2 (average 36 mm2) was placed within the solid enhancing tumor to avoid necrotic, hemorrhagic, or cystic areas that would artificially elevate or confound ADC readings. Given high inter-rater agreement, the average of both raters' measurements was used for analysis. Lower ADC values indicate restricted diffusion, a hallmark of high cellular density in lymphoma tumors.
Statistical analysis: Chi-square or Fisher's exact test was used for categorical comparisons between R/R and non-R/R groups. Mann-Whitney U test was applied for continuous variables. ROC analysis with AUC was performed for each ML model and ADC. Kaplan-Meier curves assessed progression-free survival, with log-rank tests comparing groups split by cut-off scores. A p-value below 0.05 was the threshold for statistical significance.
Of the 46 PCNSL patients, 20 (43.5%) were classified as R/R and 26 (56.5%) as non-R/R. Among the 46 patients, 27 (58.7%) achieved complete response (CR) to first-line chemotherapy, 12 (26.1%) had partial response or stable disease (PR/SD), and 7 (15.2%) exhibited progressive disease (PD). The objective response rate was 84.8% (39/46), and 20 patients (43.5%) died during the study period. The median follow-up was 27.2 months overall, and the R/R group had a median time to relapse of 13 months.
ML model comparison: All five ML model scores were significantly higher in the R/R group compared to the non-R/R group (p less than 0.05 for each). The SVM model achieved the best overall performance: accuracy of 83%, precision of 80%, and AUC of 0.78 (95% CI: 0.63-0.93, p = 0.001). The other models performed as follows: KNN at accuracy 74%, precision 70%, AUC 0.73 (0.57-0.89, p = 0.008); LDA at accuracy 72%, precision 73%, AUC 0.68 (0.51-0.84, p = 0.041); NB at accuracy 76%, precision 76%, AUC 0.72 (0.56-0.88, p = 0.012); DT at accuracy 74%, precision 70%, AUC 0.74 (0.60-0.89, p = 0.005).
ADC comparison: ADC values were significantly lower in the R/R group (median 0.67 x 10-3 mm2/s, IQR 0.62-0.72) compared to the non-R/R group (median 0.75 x 10-3 mm2/s, IQR 0.66-0.83; p = 0.017). The ADC model achieved an AUC of 0.71 (0.55-0.86), performing below the SVM (AUC 0.78), KNN (0.73), NB (0.72), and DT (0.74) models but above LDA (0.68). The SVM's superiority over ADC is notable given that SVM uses five automatically extracted radiomic features while ADC relies on manual ROI placement by expert neuroradiologists, suggesting radiomic features capture prognostically relevant information beyond what the ADC metric provides.
Clinical variables: No statistically significant differences were found between R/R and non-R/R groups for age, sex, tumor location, enhancement pattern, necrosis, hemorrhage, peritumoral edema, leptomeningeal seeding, number of lesions, maximal tumor diameter, LDH level, or Ki-67 proliferation index. This underscores that conventional clinical and visual MRI parameters cannot reliably distinguish patients who will relapse, further motivating the radiomic approach.
Beyond binary classification of R/R status, the study examined whether radiomic model scores could stratify progression-free survival (PFS). PFS was measured from the start of effective treatment to tumor progression, recurrence, or death. Kaplan-Meier analyses were conducted for all five ML models and ADC values, with patients split above and below each model's optimal cut-off score as identified by ROC analysis.
SVM and NB as survival predictors: Among the five ML models, only SVM (cut-off 0.49) and NB (cut-off 0.52) demonstrated statistically significant differences in PFS between their respective high-score and low-score patient groups (p less than 0.05 by log-rank test). Patients with SVM scores above 0.49 exhibited significantly shorter PFS than those below, confirming that the SVM score captures not just cross-sectional R/R status but also the dynamic trajectory of disease over time. KNN, LDA, and DT scores did not significantly stratify PFS in Kaplan-Meier analysis, despite their significant differences in R/R group scores.
ADC as a survival predictor: ADC values at the cut-off of 0.68 x 10-3 mm2/s did not significantly stratify PFS in the Kaplan-Meier analysis (log-rank test p greater than 0.05), even though ADC was significantly associated with R/R group membership. This dissociation suggests that while low ADC identifies tumors more likely to relapse, it does not capture the quantitative nuance needed to stratify survival trajectories, whereas SVM and NB scores based on multi-dimensional texture features are better aligned with the biological determinants of disease progression.
Practical interpretation: The finding that SVM and NB scores predict PFS while ADC does not supports the conceptual argument for radiomics: that multi-feature texture analysis extracts biologically informative signal from MRI that cannot be reduced to a single measurement. In clinical practice, a preoperative SVM score above the 0.49 threshold could potentially flag patients for intensified consolidation therapy, enrollment in clinical trials, or more frequent surveillance imaging, before any clinical evidence of relapse emerges.
Small, retrospective, two-center dataset: The most fundamental limitation is the 46-patient cohort spanning 15 years from two institutions. Although the rare incidence of PCNSL makes larger datasets difficult to assemble, 46 patients provides limited statistical power, and results from 10-fold cross-validation on such a small dataset may overestimate generalizability. The retrospective design introduces inherent selection bias: only patients with complete pretreatment and post-treatment MRI follow-up were included, which may systematically exclude patients lost to follow-up or those who declined further imaging after rapid disease progression.
No external validation: The models were evaluated only on internal cross-validation within the two-center dataset. Performance on an independent external cohort from different institutions with different MRI scanners and acquisition protocols has not been tested. The imaging equipment in this study ranged from 1.5-T to 3-T scanners across multiple manufacturers (Siemens and GE), which already introduces variability that could affect radiomic feature reproducibility. Radiomic features are known to be sensitive to acquisition parameters, reconstruction algorithms, and scanner hardware, making the generalizability of the selected five features uncertain.
Scanner and field-strength heterogeneity: Combining 1.5-T and 3-T MRI data without explicit harmonization is a recognized source of radiomic variability. Prior phantom studies found that only about one-third of MRI texture features demonstrate strong robustness across different scanner platforms. While the authors applied Z-score normalization as a harmonization step, this approach addresses intensity scale differences but does not fully correct for resolution, noise, or contrast differences between scanner types.
Limited sample for deep learning and model complexity: The authors explicitly chose simple ML algorithms with default settings over deep learning approaches (e.g., nnU-Net for segmentation, 3D-CNN for end-to-end prediction) because the dataset size was insufficient to train complex models without severe overfitting. This decision was correct given the constraints, but it means the study cannot demonstrate the ceiling of predictive performance achievable with larger datasets and more advanced architectures. Future prospective studies with multi-institutional cohorts are needed to validate and potentially improve upon these models.
The study's findings support the concept that preoperative MRI-based radiomics can provide actionable risk stratification in PCNSL before treatment initiation. In practical clinical terms, a validated high-risk SVM score at baseline could influence several treatment decisions: selection of more intensive induction regimens, early consolidation with autologous stem cell transplantation for fit patients, enrollment in clinical trials testing novel salvage therapies, and scheduling more frequent surveillance MRI intervals to detect early relapse. Currently, these decisions are made based on limited prognostic information, because conventional clinical variables (LDH, age, performance status) do not reliably predict R/R status in PCNSL.
Multi-sequence integration as a strength: One of the key methodological contributions of this study relative to prior work is the simultaneous use of three MRI sequences (CE T1WI, T2WI, and T2 FLAIR) for radiomic feature extraction. Most previous PCNSL radiomic studies relied on a single sequence. Multi-sequence integration captures complementary tissue information: CE T1WI reflects the blood-brain barrier disruption and tumor vascularity, T2WI captures tissue water content and peritumoral edema, and FLAIR highlights non-enhancing infiltrative disease components. Future studies should continue this multi-sequence approach and explore diffusion-derived metrics beyond ADC, such as mean diffusivity and fractional anisotropy from diffusion tensor imaging.
Deep learning and convolutional neural networks: The authors explicitly frame current simple ML models as a stepping stone toward deep learning approaches. As more PCNSL cases are collected in prospective multicenter registries, convolutional neural networks (CNNs) trained end-to-end on MRI volumes could learn spatial patterns invisible to hand-crafted radiomic features. 3D CNNs could capture the volumetric tumor architecture across sequences simultaneously, potentially outperforming the five-feature SVM model. Federated learning frameworks across PCNSL treatment centers could enable large-scale model training without compromising patient data privacy.
Integration with molecular and clinical data: PCNSL biology is increasingly understood at the molecular level, with DLBCL cell-of-origin subtype (GCB vs. ABC), MYD88 mutation status, and CD79B mutations known to affect prognosis and treatment response. Radiomic models that incorporate these molecular variables alongside imaging features represent the next frontier. Similarly, circulating tumor DNA (ctDNA) liquid biopsy approaches could provide dynamic prognostic updates as treatment progresses, complementing the pretreatment static MRI radiomic risk score described here.