Primary central nervous system lymphoma (PCNSL) is a rare but highly aggressive extra-nodal non-Hodgkin lymphoma confined to the brain, spine, meninges, and eyes. More than 90% of cases are histologically classified as diffuse large B-cell lymphoma (DLBCL). Without treatment, the median overall survival of PCNSL patients is only about 1.5 months, reflecting its particularly aggressive biology compared to systemic DLBCL. The disease is driven by significant genetic and molecular heterogeneity that governs both aggressiveness and tumor progression.
Role of BCL-6 as a prognostic biomarker: Among the major molecular markers associated with PCNSL prognosis, B-cell lymphoma-6 (BCL-6) has emerged as a clinically important and unfavorable marker. BCL-6 is a transcriptional repressor that plays essential roles in germinal center B-cell development and lymphomagenesis. It is the most commonly translocated gene in DLBCL and defines the germinal center B-like (GCB) molecular subtype. Multiple studies, including the prospective multicenter CALGB 50202 trial conducted in 2013, have documented a significant association between BCL-6 expression and decreased survival rates in PCNSL. Additional studies have confirmed BCL-6 as an independent unfavorable prognostic biomarker, distinguishing it from other co-expressed markers such as BCL-2 and MYC.
The clinical problem with current detection: BCL-6 status in PCNSL is currently determined through immunohistochemical staining on tissue obtained by biopsy. Pathologic diagnosis in PCNSL primarily relies on fine-needle aspiration or stereotactic biopsy, both of which are invasive procedures carrying real risks, including hemorrhage and neurological injury. These procedures are also subject to sampling error from tumor heterogeneity. Because nearly all PCNSL patients undergo brain MRI before treatment, a non-invasive MRI-based approach to detecting BCL-6 status would offer significant clinical value, allowing better upfront prognostication without additional procedural risk.
This 2025 study from the Chinese PLA General Hospital, published in Neuroradiology (European Radiology group), aims to determine whether radiomics features extracted from multiparametric MRI, combined with machine learning classifiers, can accurately detect BCL-6 overexpression status in immunocompetent PCNSL patients without tissue analysis.
This was a retrospective single-center study approved by the ethics committee of Chinese PLA General Hospital (Registration No. S2022-671-01), with the requirement for written informed consent waived. The study enrolled patients diagnosed between January 2013 and July 2023. From an initial pool of 69 pathologically confirmed PCNSL patients retrieved from the picture archiving and communication system (PACS), four were excluded due to steroid treatment before MRI or poor image quality, leaving 65 patients with a total of 101 lesions for analysis.
BCL-6 classification criteria: All patients had pathologically confirmed DLBCL by either stereotactic biopsy (58.46%), partial excision (13.85%), or whole excision (27.69%). BCL-6 status was defined by immunohistochemical staining: more than 30% staining was classified as BCL-6 overexpression (BCL-6 positive, or BCL-6+), and 30% or below as BCL-6 underexpression (BCL-6 negative, or BCL-6-). Using this threshold, 40 patients (62 lesions) were BCL-6+ and 25 patients (39 lesions) were BCL-6-. The BCL-6+ group was 57.5% male with a mean age of 58.11 years, while the BCL-6- group was 64% male with a mean age of 56.59 years. No significant differences in age (P = 0.36), sex (P = 0.60), or number of lesions (P = 1.00) existed between groups.
MRI acquisition: All scans were performed on a 3T GE Healthcare Discovery 750 MRI system with a 32-channel head coil. Three sequences were used: axial fast spin echo T2-weighted imaging (T2WI, TR/TE = 5642/93 ms, FOV = 24x24 cm, matrix = 512x512, NEX = 1.50), coronal T2 fluid-attenuated inversion recovery (T2FLAIR, TR/TE/TI = 8527/162/2100 ms, matrix = 288x224, NEX = 1.00), and axial diffusion-weighted imaging (DWI, TR/TE = 3000/65.5 ms, b = 0/1000 s/mm2, matrix = 160x160, NEX = 2.00). Apparent diffusion coefficient (ADC) maps were derived from the DWI data. All sequences used a slice thickness of 5.0 mm with a 1.5 mm gap.
Train/validation split: The 101 lesions were randomly divided into a training set (81 lesions: 49 BCL-6+ and 32 BCL-6-) and a validation set (20 lesions: 13 BCL-6+ and 7 BCL-6-) in an 8:2 ratio. Notably, only non-contrast sequences were used, making this approach applicable to patients who cannot receive gadolinium-based contrast agents.
Tumor segmentation was performed on DICOM images converted to NIfTI format using ITK-SNAP software. The regions of interest (ROI) included solid tumor components and necrosis or cystic changes, but excluded hemorrhage. To ensure reliable measurements, only lesions with a minimum diameter greater than 5 mm were included. Three-dimensional segmentations were independently delineated on ADC, T2WI, and T2FLAIR by two radiologists each with six years of experience. Interobserver consistency was quantified using the intraclass correlation coefficient (ICC), with ICC >= 0.80 required to indicate strong reliability. Features passing this threshold were retained for further analysis.
Feature extraction: Radiomic features were extracted using a customized version of PyRadiomics (Python 3.8.8). From each tumor segment in each MRI sequence, a total of 2,234 radiomic characteristics were extracted. These features spanned three categories: 14 shape features, 36 first-order histogram features, and 75 second-order texture features (also called grey-level co-occurrence matrix and related features). Feature extraction was further expanded by applying multiple image filters to the original images, including 8 wavelet transforms, 5 log-sigma Laplacian of Gaussian transforms, square, square root, logarithmic, exponential, gradient, and 2D LBP transforms, generating a comprehensive feature space across multiple spatial frequency bands.
Feature selection pipeline: Feature reduction was applied in four steps. First, only features with ICC >= 0.80 were retained, ensuring reproducible measurements between segmenters. Second, Levene's test assessed distributional properties, and a t-test was applied to assess correlation between each feature and BCL-6 expression status, retaining features with P less than 0.05. Third, least absolute shrinkage and selection operator (LASSO) regression was applied to further reduce the feature set by penalizing coefficients and driving uninformative features to zero. Fourth, the final feature subset was filtered based on feature weight rankings. After this four-step pipeline, the number of retained features per sequence group ranged from 4 features for T2WI alone to 17 features for the combined ADC+T2WI+T2FLAIR group.
Sequence groupings: Six distinct sequence groups were evaluated: ADC alone, T2WI alone, T2FLAIR alone, ADC+T2WI, ADC+T2FLAIR, and the combined triple-sequence group ADC+T2WI+T2FLAIR. One potential combination (T2WI+T2FLAIR) was excluded because no features passed the selection criteria, leaving five meaningful single/dual-sequence groups plus the triple-sequence group for a total of six groups tested against five classifiers, generating 30 distinct models.
The final selected features from each sequence group were fed into five machine learning classifiers built on the scikit-learn Python toolkit: logistic regression (LR), Naive Bayes (NB), support vector machine (SVM), K-nearest neighbor (KNN), and multilayer perceptron (MLP). Each classifier represents a different approach to decision boundary construction. LR fits a linear boundary; NB applies Bayesian probability estimation assuming feature independence; SVM constructs an optimal separating hyperplane using the principles of VC dimension and minimum structural risk theory; KNN classifies based on the labels of the nearest training samples in feature space; and MLP uses a fully connected neural network with one or more hidden layers to learn non-linear mappings.
Cross-validation and evaluation metrics: All models were trained and internally validated using five-fold cross-validation on the 81-lesion training set. Default hyperparameters from scikit-learn were used for all classifiers without additional tuning, which is an important methodological detail because it means the results represent out-of-the-box performance rather than heavily optimized configurations. Model performance was evaluated across five metrics: sensitivity, specificity, accuracy, F1-score (the harmonic mean of precision and recall), and area under the receiver operating characteristic curve (AUC), providing a comprehensive view of each model's discriminatory ability across different decision thresholds.
Tumor morphology assessment: In addition to radiomic features, conventional MRI morphologic features were assessed and compared between BCL-6+ and BCL-6- groups. These included tumor location (supratentorial vs. infratentorial, midline vs. peripheral), maximum and minimum diameters, ADC quantitative parameters (ADCmean, ADCmax, ADCmin, ADC 5th percentile, and ADC uniformity), presence of necrosis or cystic change, edema grade, and mass effect level. These assessments were performed by three radiologists (two with six years of experience, one with three years) to characterize what conventional imaging alone can and cannot detect.
The combination of sequence groups and classifiers generated 30 distinct model configurations, allowing a systematic comparison of how both the choice of input data and the classification algorithm independently and jointly affect BCL-6 detection performance in this dataset.
Among all 30 models evaluated, the SVM classifier combined with the triple-sequence group (ADC+T2WI+T2FLAIR) delivered the strongest overall performance. In the training set, this model achieved an AUC of 0.95 (95% CI: 0.883-1.000), sensitivity of 92.0%, specificity of 88.0%, accuracy of 93.0%, and F1-score of 0.94. In the independent validation set of 20 lesions, the same SVM triple-sequence model achieved an AUC of 0.87 (95% CI: 0.703-1.000), sensitivity of 98.0%, specificity of 84.0%, and accuracy of 90.0%, demonstrating consistent generalization from training to validation.
Single-sequence model performance: Among the single-sequence models in the training set, the ADC-SVM model performed best with AUC = 0.85 (sensitivity 92%, specificity 75%, accuracy 88%). The T2WI-KNN model reached AUC = 0.84, while T2FLAIR-based models generally performed below 0.80 across classifiers. This hierarchy suggests that ADC carries the most discriminative information among the individual sequences, consistent with the known association between diffusion restriction and cellular density in DLBCL. In the validation set, ADC-SVM maintained AUC = 0.89 (sensitivity 92%, specificity 83%, accuracy 86%), performing notably well as a single-sequence model.
MLP performance: The MLP classifier also showed competitive results with the triple-sequence group, achieving AUC = 0.93 in training (sensitivity 75%, specificity 88%, accuracy 86%, F1 = 0.89). In the validation set, MLP reached AUC = 0.80 with the triple-sequence group, somewhat below the SVM. Across all classifiers, adding sequences consistently improved performance for most models, confirming the complementary information contributed by T2WI texture features and T2FLAIR signal characteristics beyond what ADC diffusion maps alone provide.
Conventional MRI morphology findings: None of the conventional morphological features showed significant differences between BCL-6+ and BCL-6- groups, including location (P greater than 0.05), maximum diameter (P = 0.15), minimum diameter (P = 0.35), ADCmax (P = 0.76), ADCmin (P = 0.10), ADC 5th percentile (P = 0.18), ADC uniformity (P = 0.26), necrosis or cystic change (P = 0.13), edema (P = 0.56), or mass effect (P = 0.81). The one exception was ADCmean, which showed a statistically significant difference (P = 0.00): BCL-6+ lesions had lower mean ADC values (0.83 x 10-3 mm2/s) compared to BCL-6- lesions (0.95 x 10-3 mm2/s). The IELSG prognostic score also differed significantly between groups (P = 0.02).
The study situates its findings within the broader biology of DLBCL and PCNSL. DLBCL exhibits two major molecular subtypes defined by gene expression patterns: germinal center B-like DLBCL (GCB) and activated B-like DLBCL (ABC). The GCB subtype prominently features the BCL-6 gene as a germinal center marker. BCL-6 protein functions as a transcriptional repressor that modulates lymphocyte differentiation and promotes lymphomagenesis through downstream target gene regulation. In PCNSL specifically, BCL-6 expression was shown to correlate with decreased survival in the CALGB 50202 prospective trial, and multiple subsequent studies have confirmed BCL-6 as a clinically relevant unfavorable biomarker in both immunocompetent and immunocompromised settings.
Interpretation of ADCmean findings: The significant difference in ADCmean between BCL-6+ and BCL-6- lesions (P = 0.00, with lower values in BCL-6+ tumors) is biologically coherent. ADC values quantify the diffusion of water molecules and are inversely related to cellular density and extracellular space tortuosity. Tumors with higher cellular density, typically reflecting more aggressive biology, exhibit restricted diffusion and lower ADC values. Low ADC values have previously been identified as independent unfavorable prognostic factors in PCNSL patients, and BCL-6+ tumors with their more aggressive behavior would be expected to display tighter cellular packing. This finding positions ADCmean as a potentially useful simple screening parameter, though insufficient on its own for reliable classification.
Why radiomics outperforms conventional features: While ADCmean provides a useful signal, the full radiomic feature set captures textural and higher-order spatial information that reflects intratumoral heterogeneity in ways that single quantitative parameters cannot. Radiomics extracts geometry, intensity, and texture attributes across multiple spatial scales (via wavelet and LoG filters), effectively characterizing the microstructural patterns within tumors. The combination of ADC texture features with T2WI signal characteristics and T2FLAIR signal suppression patterns provides complementary information about cell membrane composition, water content, and protein concentration within the tumor microenvironment. This multi-sequence approach mirrors previous work by Liu et al. (2024) who used multiparametric MRI radiomics to differentiate double-expression (DEL) status in PCNSL, and Bathla et al. (2021) who applied radiomics to distinguish PCNSL from glioblastoma with an AUC of 0.87-0.94.
Practical significance of non-contrast sequences: A particularly noteworthy aspect of this study is that high diagnostic performance was achieved using only three non-contrast MRI sequences (ADC, T2WI, T2FLAIR), without post-contrast T1 enhancement or MR perfusion imaging. This is clinically important because contrast agents are contraindicated in patients with renal insufficiency and in those with gadolinium deposition concerns. The ability to detect BCL-6 status from non-contrast sequences means the approach is applicable to a broader patient population, including elderly PCNSL patients who frequently have comorbid kidney disease.
Single-center design with a single scanner: The most significant limitation is that all 65 patients were imaged on a single GE Healthcare Discovery 750 3T scanner at one institution over a ten-year period. Single-center radiomics studies are inherently subject to selection bias and may not generalize to other MRI platforms, acquisition protocols, field strengths, or patient populations. Radiomic features are known to be sensitive to scanner hardware differences, acquisition parameters, and image reconstruction settings, and features that are discriminative at one institution may not remain so when the model is applied to data from different scanners. External multi-institutional validation is essential before this approach can be considered for clinical deployment.
Small sample size and class imbalance: With only 65 patients and 101 lesions total, the dataset is relatively small for machine learning model development, particularly given the 8:2 train/validation split that leaves only 20 lesions in the validation set. The class imbalance between BCL-6+ (62 lesions) and BCL-6- (39 lesions) in the training set introduces potential bias toward the majority class. Five-fold cross-validation within the training set provides some mitigation, but cannot substitute for a larger and more balanced dataset or a fully independent external cohort.
Absence of post-contrast and perfusion sequences: The study deliberately excluded post-contrast T1 imaging and MR perfusion techniques (such as dynamic contrast-enhanced MRI and dynamic susceptibility contrast MRI). While this is a practical strength for broader applicability, it also means the models were not tested with the full range of available MRI biomarkers. Post-contrast enhancement patterns and perfusion parameters (such as cerebral blood volume) are known to correlate with tumor vascularity and blood-brain barrier disruption in PCNSL and may carry additional discriminative information for BCL-6 status. Future studies should assess whether incorporating these sequences further improves classification performance.
Scope limited to BCL-6 only: This study focused exclusively on BCL-6 expression, leaving other clinically relevant molecular markers in PCNSL unaddressed, including Ki-67 (proliferation index), BCL-2, and MYC. In practice, PCNSL prognosis is influenced by the interplay between these markers, and single-marker detection may capture only part of the molecular prognostic landscape. Additionally, all machine learning classifiers were run with default scikit-learn parameters without hyperparameter optimization, which may mean that the reported performance represents a lower bound on achievable accuracy rather than a fully optimized model.
Multi-institutional external validation: The immediate next step for this work is external validation across multiple centers with different MRI platforms. Given that PCNSL is a rare disease, multi-institutional collaboration is particularly important for assembling datasets large enough to robustly evaluate model performance and failure modes. Federated learning approaches, which allow model training across institutions without sharing raw patient data, offer a practical path to building larger training sets while respecting data governance constraints. Such validation studies would also need to address scanner harmonization and feature normalization to ensure that radiomic models trained at one site maintain performance at others.
Multi-marker molecular detection: BCL-6 is one of several prognostically relevant molecular markers in PCNSL. Future work should extend the radiomic machine learning framework to simultaneously or sequentially detect BCL-2, MYC, and Ki-67 expression status. Simultaneous multi-marker prediction from a single MRI acquisition would be particularly valuable, as it could replace an entire immunohistochemistry panel for initial molecular profiling. The prior study by Liu et al. (2024) demonstrating that multiparametric MRI radiomics can distinguish double-expression (BCL-2 and MYC co-overexpression) status supports the feasibility of this direction.
Incorporating contrast and perfusion sequences: Adding post-contrast T1 enhancement characteristics, dynamic contrast-enhanced (DCE) MRI perfusion parameters, and DSC-MRI blood volume estimates to the radiomic feature set could substantially improve classification performance beyond what non-contrast sequences alone provide. The blood-brain barrier disruption patterns visible on post-contrast imaging reflect vascular biology that may correlate with BCL-6-driven transcriptional programs, and perfusion parameters have previously been shown to predict outcomes in PCNSL independently. The authors explicitly call for future studies to evaluate these modalities.
Hyperparameter optimization and deep learning: The current study used default scikit-learn parameters for all five classifiers, intentionally keeping the comparison methodologically simple. Future work should apply systematic hyperparameter tuning, including grid search or Bayesian optimization, to determine whether further performance gains are achievable. Separately, end-to-end deep learning approaches (such as 3D CNNs operating directly on volumetric MRI data without handcrafted radiomic features) represent a complementary direction that avoids the assumptions embedded in the radiomics pipeline and may capture additional spatial patterns predictive of BCL-6 status.