Primary central nervous system lymphoma (PCNSL) is a rare, highly aggressive extranodal non-Hodgkin lymphoma confined to the brain, spinal cord, leptomeninges, and eyes. It comprises only approximately 4% of newly diagnosed CNS tumors and 4-6% of all extranodal lymphomas in immunocompetent patients. Its annual incidence in the United States ranges from 0.3 to 0.6 cases per million, yet this rate has risen steadily over the past four decades, particularly in patients over 60. The 5-year overall survival (OS) rate ranges from just 22.3% to 35%, reflecting the disease's clinical difficulty despite therapeutic progress.
Histopathological profile: Approximately 95% of PCNSL cases are classified as diffuse large B-cell lymphoma (DLBCL) with a predominantly non-germinal center B-cell-like (non-GCB) immunophenotype. Characteristic histological features include a perivascular arrangement of highly proliferating tumor cells forming a distinctive angiocentric growth pattern. Despite these recognizable features, 15-25% of patients fail to respond to chemotherapy and 25-50% relapse after an initial response, leaving significant clinical gaps.
Limitations of existing prognostic models: Two widely used prognostic scoring systems exist: the International Extranodal Lymphoma Study Group (IELSG) model, which incorporates age, performance status, serum lactate dehydrogenase (LDH), cerebrospinal fluid (CSF) protein, and deep brain involvement; and the Memorial Sloan-Kettering Cancer Center (MSKCC) model, which uses only age and performance status. Both have significant limitations. LDH and CSF protein values are not always available at diagnosis, and lumbar puncture carries contraindications in some patients. More critically, several recent studies have failed to validate the MSKCC model's ability to identify significant survival differences. New prognostic biomarkers are urgently needed.
This 2024 study from Beijing Tiantan Hospital, published in the Journal of Neuro-Oncology, addresses this gap by developing and validating a novel machine learning-based pathomics signature derived entirely from routine hematoxylin and eosin (H&E)-stained histopathology slides, without requiring additional molecular testing or specialized assays.
The study enrolled 114 patients with histologically confirmed CNS-DLBCL from Beijing Tiantan Hospital, spanning January 2019 to March 2023. A total of 132 whole-slide images (WSIs) were collected across these patients; some contributed more than one slide. Patients were divided into two cohorts: a training cohort (cohort 1) of 68 patients with 71 WSIs used to build the prognostic model, and an independent validation cohort (cohort 2) of 46 patients with 61 WSIs used to test it. This explicit train/validate split is methodologically important in a rare disease context where data from public repositories like TCGA are unavailable.
Inclusion and exclusion criteria: Inclusion required histological diagnosis of CNS-DLBCL, absence of concomitant tumors, and availability of complete clinicopathological and follow-up data. Patients were excluded if systemic DLBCL was evident on CT or PET-CT of the chest, abdomen, pelvis, or bone marrow aspiration; if WSIs were incomplete or low quality; or if clinical or follow-up data were missing. This rigorous screening ensured a clean, PCNSL-specific dataset.
Patient characteristics: Among the 114 patients, 57 (50%) were male, with a median age of 64 years (IQR 54-69). Stereotactic biopsy was the most common diagnostic procedure (89.47%), with only 12 patients (10.53%) undergoing surgical resection or open biopsy. Based on the Hans algorithm, 71.93% were classified as GCB subtype and 28.07% as non-GCB. Deep brain involvement was present in 74.56% and multifocal lesions in 58.77%. Approximately 36.84% received BTK inhibitor therapy and 28.07% received consolidation therapy. Median OS for the combined cohort was 34.07 months (95% CI: 24.50 to not reached) and median PFS was 12.70 months (95% CI: 9.20-23.30).
All slides were formalin-fixed, paraffin-embedded, and scanned at 20x magnification using a Leica Aperio CS2 scanner into standard SVS format. Two experienced pathologists annotated regions of interest (ROI) using ASAP version 2.1, with a third senior pathologist adjudicating discordant annotations, ensuring annotation quality appropriate for computational analysis.
The core methodological innovation is a fully automated pipeline built on CellProfiler (version 4.2.6), an open-source image analysis software widely used in high-throughput cellular phenotyping. After ROI annotation, slides were divided into 512x512 pixel patches using OpenSlide, and color normalization was applied via the Vahadane method to minimize staining variability across slides. To reduce computational load while retaining representativeness, 50 non-overlapping patches with high tumor cell density were selected from each patient by experienced pathologists for feature extraction.
Feature extraction pipeline: CellProfiler decomposed each image into separate hematoxylin-stained and eosin-stained grayscale channels using the UnmixColors module. Tumor cell nuclei were identified with the IdentifyPrimaryObjects module, cytoplasm was computed by subtracting nuclei from the full cell body (IdentifyTertiaryObjects), and cell bodies were identified with the IdentifySecondaryObjects module. From these compartments, eight measurement modules were applied: MeasureImageQuality, MeasureImageIntensity, MeasureGranularity, MeasureColocalization, MeasureObjectIntensity, MeasureObjectNeighbors, MeasureObjectSizeShape, and MeasureTexture. Averaging feature values across all 50 patches per patient yielded a final 802-feature matrix per subject.
Diagnostic classification step: Before building the prognostic model, the team first validated that these 802 features could biologically distinguish tumor from normal adjacent tissue. Six machine learning classifiers were tested: Logistic Regression, K-Nearest Neighbor (KNN), Random Forest (RF), Support Vector Machine (SVM), XGBoost, and Decision Tree (DT). The dataset was split 60% training / 40% testing. All classifiers achieved AUC above 0.9 on the training set. On the test set, the average AUC was 0.968, with individual classifiers ranging from Decision Tree (AUC 0.891) to SVM and RF (AUC 0.990 and 0.994, respectively). This strong discriminative performance validated that the extracted features carry meaningful biological signal.
The top discriminative features selected by Random Forest included Image Granularity, Image Colocalization, Haralick texture features, Image Quality metrics, and Object Intensity features, establishing a basis for the subsequent prognostic feature selection step.
Building the prognostic pathomics score required a three-step feature selection procedure applied only to the training cohort, guarding against overfitting to the combined dataset. First, univariate Cox regression analysis screened all 802 features, retaining only those with p less than 0.05 as candidate prognostic features. This yielded 91 candidate features. Second, LASSO-Cox (Least Absolute Shrinkage and Selection Operator combined with Cox proportional hazards regression) was applied with tenfold cross-validation to determine the optimal L1 regularization parameter lambda. LASSO shrinks coefficients of weakly predictive features to exactly zero, providing both feature selection and regularization in one step. The optimal lambda of 0.186618 (log-lambda = -1.679) selected 22 features with non-zero coefficients.
Final model construction: A backward stepwise multivariate Cox proportional hazards model further refined these 22 features to a final eight-feature Path-score formula. Proportional hazard assumptions were confirmed via the Schoenfeld Individual Test (Global Schoenfeld Test p = 0.230, indicating no violation). The final Path-score formula is a linear combination of eight weighted features: Granularity_3_Hematoxylin (coefficient 2.341), Mean Cells AreaShape Zernike 7,1 (0.015), Mean Cytoplasm AreaShape Zernike 4,2 (5.140), Mean Cytoplasm Texture SumVariance Hematoxylin 3,00,256 (0.297), Mean Nuclei AreaShape Zernike 1,1 (4.692), Mean Nuclei Intensity MaxIntensity Hematoxylin (3.349), Mean Nuclei Intensity MinIntensityEdge Hematoxylin (0.042), and Texture SumVariance Hematoxylin 3,01,256 (3.387).
Biological interpretation of selected features: Zernike shape features decompose the circular projection of each nucleus or cytoplasm into polynomial coefficients, capturing subtle asymmetries and shape complexity that correlate with nuclear pleomorphism, a hallmark of aggressive tumor biology. Texture features such as SumVariance measure correlation among neighboring pixel intensities within defined regions, quantifying intratumoral heterogeneity at the microscopic scale. Nuclei intensity features reflect chromatin density and distribution, which are known to correlate with proliferative activity. Together, these eight features capture local cell morphology (shape of nucleus, cell, cytoplasm) and global tissue patterns (texture), both of which were independently associated with survival in the multivariate analysis.
The Path-score for the validation cohort was computed using the exact formula derived from the training cohort, without any re-training. Patients were stratified into high-risk and low-risk groups using the optimal cutoff of 1.824 determined by maximally selected rank statistics in the training cohort, and the same threshold was applied unchanged to the validation cohort.
Kaplan-Meier survival analysis revealed dramatically different outcomes between high- and low-Path-score patients in the training cohort. Patients with high Path-score had median OS of only 11.10 months (95% CI: 6.17 to not reached), compared to not reached (95% CI: NR to NR) for low Path-score patients (log-rank p less than 0.001). For PFS, high Path-score patients had median PFS of 5.72 months (95% CI: 2.00 to NR) versus 16.80 months (95% CI: 9.80 to NR) for low Path-score patients (log-rank p = 0.001).
Independent validation cohort results: The same stratification pattern held in the independent validation cohort. High Path-score patients had median OS of 23.80 months (95% CI: 11.3 to NR) and median PFS of 5.93 months (95% CI: 4.47 to NR), while low Path-score patients showed median OS of not reached (95% CI: 25.60 to NR) and median PFS of 29.43 months (95% CI: 12.00 to NR). Log-rank p-values were 0.016 for OS and 0.001 for PFS, confirming the signature's generalizability beyond the training dataset.
Multivariate Cox regression: After adjusting for clinicopathological variables in the training cohort, high Path-score remained an independent predictor of OS (HR 7.423, 95% CI: 2.738-20.119, p less than 0.001) and PFS (HR 2.143, 95% CI: 1.116-4.113, p = 0.022). In the validation cohort multivariate analysis, the Path-score retained independent prognostic value for OS (HR 4.204, 95% CI: 1.299-13.601, p = 0.017) and PFS (HR 3.243, 95% CI: 1.440-7.301, p = 0.005). These hazard ratios are large by oncology standards, particularly the 7.4-fold OS hazard in the training cohort.
Subgroup forest plot analysis confirmed that the Path-score remained an effective predictor (HR greater than 1, p less than 0.05) across multiple clinical subgroups, indicating the signature's robustness across different patient subsets rather than being driven by a single subgroup's characteristics.
Beyond survival prediction, the researchers investigated whether the Path-score could predict response to initial treatment, which is clinically important because identifying likely non-responders before or at diagnosis could guide more aggressive upfront therapy. In the combined cohort of 105 patients with available response data, patients with progressive disease (PD) as best response had significantly higher Path-scores (median 2.313, IQR 1.052-5.467) compared to patients achieving complete remission (CR; median 0.858, IQR 0.396-1.824), partial remission (PR; median 0.775, IQR 0.462-0.875), or stable disease (SD; median 0.859, IQR 0.719-1.014). The difference across response categories was highly significant (Kruskal-Wallis p = 0.007).
Responders vs. Non-Responders: When analyzed as a binary classification (Responders defined as CR or PR; Non-Responders as SD or PD), patients who responded to treatment had significantly lower Path-scores than Non-Responders (median 0.849, IQR 0.403-1.743 vs. median 2.150, IQR 0.942-4.833; Wilcoxon p = 0.002). Among patients in the high Path-score group, 54.29% (19/35) were Non-Responders, compared to only 22.86% (16/70) in the low Path-score group (p less than 0.001). This represents a 2.4-fold higher non-response rate in the high-risk group.
Primary resistance: High Path-score was strongly associated with primary tumor resistance, defined as progression during the initial treatment course. The primary resistance rate in the high Path-score group was 54.29% (19/35 patients) versus 18.57% (13/70 patients) in the low Path-score group (p less than 0.001). The AUC of the Path-score for predicting treatment response was 0.684 (95% CI: 0.574-0.794), for predicting primary resistance was 0.706 (95% CI: 0.593-0.819), and for predicting recurrence was 0.618 (95% CI: 0.505-0.732).
Patients with disease recurrence also showed higher Path-scores than those without recurrence (1.337, IQR 0.699-4.063 vs. 0.884, IQR 0.393-1.823; p = 0.042). These results collectively indicate that the pathomics features captured by the pipeline reflect intratumoral heterogeneity patterns that determine both survival outcomes and therapeutic responsiveness, potentially serving as a pre-treatment biomarker to guide clinical decision-making.
To produce a clinically deployable individualized prediction tool, the Path-score was integrated with independent clinical prognostic factors identified by backward stepwise multivariate Cox regression. In addition to the Path-score, the Karnofsky Performance Status (KPS) and biopsy type were independently associated with OS and were incorporated into a three-variable nomogram predicting 1-year, 2-year, and 3-year overall survival. The C-index of the nomogram in the training cohort was 0.849 (95% CI: 0.790-0.908) and in the validation cohort was 0.747 (95% CI: 0.608-0.886), representing a significant improvement over the Path-score alone (training C-index: 0.745, p = 0.001; validation C-index: 0.623, p = 0.009).
Time-dependent AUC performance: The nomogram achieved time-dependent ROC AUC values of 0.862 (95% CI: 0.772-0.953), 0.932 (95% CI: 0.835-1.000), and 0.927 (95% CI: 0.787-1.000) at 1, 2, and 3 years in the training cohort, respectively. In the validation cohort, AUC values were 0.802 (95% CI: 0.624-0.980), 0.768 (95% CI: 0.576-0.960), and 0.938 (95% CI: 0.837-1.000) at 1, 2, and 3 years. Calibration curves showed favorable agreement between nomogram-predicted and actual survival in both cohorts, confirming the model is well-calibrated rather than merely discriminative.
Head-to-head comparison with IELSG and MSKCC: The nomogram demonstrated significantly better discrimination than both existing clinical models across all time points. In the training cohort, the IELSG model achieved 1-year AUC of 0.620 (p = 0.046 vs. nomogram), 2-year AUC of 0.769 (p less than 0.001), and 3-year AUC of 0.733 (p = 0.001). The MSKCC model performed similarly or worse than IELSG. Decision curve analysis (DCA) confirmed the nomogram had higher net clinical benefit at most threshold probabilities compared to Path-score alone, KPS, IELSG, and MSKCC in both cohorts.
Net reclassification improvement: The nomogram showed an NRI of 0.469 (95% CI: 0.157-0.643; p = 0.004) and an IDI of 0.152 (95% CI: 0.059-0.308; p less than 0.001) compared to the Path-score alone in the training cohort. In the validation cohort, NRI was 0.457 (95% CI: 0.082-0.767; p = 0.016) and IDI was 0.229 (95% CI: 0.047-0.501; p = 0.004). These metrics confirm that the nomogram provides meaningful incremental reclassification over any component alone.
Retrospective design and selection bias: The study's most significant limitation is its retrospective, single-center nature. Retrospective designs are susceptible to inherent selection biases and unknown confounders that cannot be fully controlled. Although the authors verified key results in an independent validation cohort drawn from the same institution, cohort 1 and cohort 2 share the same clinical environment, scanner hardware, staining protocols, and pathologist practices. This within-institution validation, while stronger than pure internal cross-validation, does not fully simulate the variability that would be encountered in an external multi-institutional cohort.
Small sample size: With only 114 patients across both cohorts, the study is limited by the intrinsic rarity of PCNSL. The incidence of roughly 0.3-0.6 per million per year makes large-scale data collection at any single institution extremely difficult, and no public PCNSL-specific imaging-genomic database currently exists. The authors acknowledge this constraint directly and note that active efforts to expand the cohort are ongoing. Models trained on small datasets may overfit to institution-specific characteristics and require substantially larger multicenter datasets before clinical deployment can be considered.
Deep learning vs. explainable pathomics: The authors deliberately chose CellProfiler-based handcrafted feature extraction over end-to-end deep learning, arguing that interpretability and reproducibility are critical for rare disease contexts where training data is scarce. Deep learning survival models directly operating on WSIs (such as those based on attention-based MIL or transformer architectures) have shown promise in other cancers but have not yet been widely applied to PCNSL. The authors acknowledge deep learning as a potential future direction, particularly as slide repositories grow and foundation models pre-trained on large pathology datasets become available for fine-tuning.
Future directions: The authors envision several extensions of this work. Integration of genomic features, radiomics from MRI, and clinical variables with the pathomics signature into a multimodal prognostic framework could further improve prediction accuracy. Prospective multicenter validation studies incorporating diverse patient populations, different staining platforms, and scanner hardware are needed before this signature could be adopted in routine clinical practice. Ultimately, combining pathomics, radiomics, and genomics into a unified model may enable precision medicine approaches to PCNSL treatment selection, including identification of patients who would benefit from early treatment intensification or clinical trial enrollment based on their predicted likelihood of primary resistance.