Primary central nervous system lymphoma (PCNSL) is a rare but highly malignant brain tumor, classified within diffuse large B-cell lymphoma, whose incidence is rising continuously, particularly in older patients. Despite its apparent homogeneity on imaging, PCNSL is biologically complex: recent multiomics analyses have identified four distinct molecular subtypes with different clinical behaviors and survival characteristics. The oncologic management landscape is also rapidly evolving, with targeted therapies and CD19-directed chimeric antigen receptor (CAR) T-cell therapies entering clinical practice alongside conventional high-dose methotrexate-based regimens.
Why tumor burden measurement matters: Per the International Primary CNS Lymphoma Collaborative Group (IPCG) consensus, lymphoma tumor burden is defined by abnormal contrast enhancement visible on gadolinium-enhanced T1-weighted MRI, reflecting disruption of the blood-brain barrier. Accurate, objective, and reproducible quantification of this enhancing tumor burden is essential for staging, treatment response assessment per the Deauville-equivalent criteria, and enrollment in clinical trials. Despite this, no validated open-source segmentation tool for PCNSL existed before this study, leaving clinicians dependent on manual delineation by neuroradiologists.
The gap this study addresses: Prior automated volumetric approaches for brain lymphoma either lacked independent external validation or were trained on glioma data and then applied to lymphoma without disease-specific retraining. When the authors tested a glioma-trained nnU-Net model on their PCNSL cohort, it performed poorly, with Dice scores of 0.14 (internal) and 0.07 (external) and mean average surface distance (MASD) values exceeding 40 mm, confirming that cross-disease transfer without retraining is unreliable for this specific tumor type.
The study trained and externally validated a dedicated deep learning segmentation model for PCNSL using 135 patients from eight French clinical centers, making this the first publicly available, externally validated automatic segmentation tool specifically designed for brain lymphoma. The model is available at a public GitHub repository to encourage widespread use and refinement by the research community.
The study was a multicenter retrospective analysis drawing from the French Network for Oculocerebral Lymphoma (LOC Network) database. Patients required pathologically confirmed, Epstein-Barr virus-negative, immunocompetent PCNSL and a gadolinium-enhanced T1-weighted MRI available for analysis. All patients were HIV-negative with no systemic lymphoma involvement. MRI scans were acquired between September 2010 and February 2022 as part of routine clinical care, meaning images were not harmonized across sites and reflect genuine real-world variability.
Internal and external cohorts: Eighty-seven patients from Pitié Salpêtrière University Hospital (Paris) formed the internal dataset, split into 43 patients for training and validation and 44 for internal testing. Forty-eight additional patients from seven other French centers (Clermont-Ferrand, Amiens, Nancy, Timone, Pasteur, Bergonié Institute, and Oncologic Institute of the West, with 6-7 patients per center) formed the external test set. Ages ranged from 32 to 95 years across the full cohort, with mean ages of 67.0 years (internal) and 75.5 years (external), reflecting the older demographic of immunocompetent PCNSL.
MRI heterogeneity: The internal dataset included images from three different field strengths (1.0 T, 1.5 T, 3 T), three manufacturers (GE HealthCare, Philips, Siemens), both 2D and 3D acquisitions, and both spin-echo and gradient-echo sequences. In-plane resolution ranged from 0.47 mm to 2 mm, and slice thickness ranged from 0.47 mm to 6.5 mm. The external dataset showed similar heterogeneity, including Toshiba scanners in addition to the three manufacturers represented internally. This hardware diversity is a deliberate feature of the study design: a model that works only on homogeneous, harmonized data would have limited clinical value.
Lesion complexity: The mean number of lesions per scan was 3.68 (SD 3.22) internally and 5.42 (SD 5.54) externally, with some patients presenting with 10 or more discrete enhancing foci. PCNSL imaging patterns were classified into five groups: unique homogeneous lesion (group A), multiple discrete homogeneous lesions (group B), single heterogeneous lesion (group C), multiple heterogeneous lesions including linear ependymal enhancement (group D), and linear perivascular enhancement (group E). This pattern classification was used to analyze subgroup performance and identify where the model struggled.
All MRI volumes were annotated by a single certified neuroradiologist (L.N.) with 5 years of neuro-oncology imaging experience using ITK-SNAP software (version 3.6) with a semiautomatic active contour method followed by manual correction. Annotations focused exclusively on contrast-enhancing tumor burden, consistent with IPCG criteria. To establish the reliability of these reference-standard annotations, both intrareader variability (the primary annotator re-segmented 50 randomly selected cases after a 1-month washout interval) and interreader variability (a second board-certified neuroradiologist with 1 year of neuro-oncology experience independently annotated the same 50 cases) were quantified.
Annotation reliability: Intrareader metrics were: Dice score 0.85 (95% CI: 0.82-0.88), F1 score 75% (95% CI: 68-82%), and MASD 1.05 mm (95% CI: 0.77-1.39 mm). Interreader metrics were slightly lower, as expected: Dice score 0.81 (95% CI: 0.78-0.85), MASD 1.64 mm (95% CI: 1.20-2.17 mm), F1 score 74% (95% CI: 68-80%). These figures are important context for interpreting model performance, since they represent the ceiling of what agreement between human experts achieves on this task.
nnU-Net V2 framework: The segmentation model used nnU-Net V2, a self-configuring deep learning framework that has consistently ranked among the top-performing algorithms across diverse medical segmentation benchmarks. The pipeline autonomously extracts dataset fingerprints (intensity distributions, image sizes, modalities) and uses rule-based heuristics to configure preprocessing, architecture, batch size, learning rate schedules, and data augmentation strategies without manual tuning. All images were resampled to isotropic 1x1x1 mm voxels and rescaled to intensity range 0-1. Fivefold cross-validation was run on the training and validation set for 1000 epochs each to select the optimal configuration. No modifications were made to the nnU-Net algorithm, preserving reproducibility.
Evaluation metrics: Following Metrics Reloaded consensus guidelines, three complementary metrics were reported. The 3D Dice score and MASD measure voxel-level segmentation accuracy at the patient level. The instance-level F1 score evaluates how well the model detects individual lesions, computed by breaking predicted and reference segmentation maps into connected components via connected-component analysis, using intersection over union (IoU) greater than 0.1 as the localization criterion, and then computing the harmonic mean of precision and recall across lesion instances per patient. Pearson correlation between automatic and manual volumetric measurements was also assessed for its clinical relevance to tumor burden quantification.
The nnU-Net model achieved a mean Dice score of 0.84 (95% CI: 0.79-0.88) on the internal test set (n = 44) and 0.88 (95% CI: 0.84-0.91) on the external test set (n = 48). The Mann-Whitney U test showed no statistically significant difference in performance between internal and external datasets (P = .59), confirming that the model generalizes across acquisition centers, scanners, and protocols without performance degradation. These Dice scores are directly comparable to, or slightly better than, the intrareader (0.85) and interreader (0.81) agreement among human experts, indicating that the model operates at a level approaching expert consensus.
MASD results: Mean MASD was 1.69 mm (95% CI: 0.94-2.67 mm) on the internal test set and 1.44 mm (95% CI: 0.90-2.21 mm) on the external test set. These values compare favorably to the interreader MASD of 1.64 mm, again suggesting that automatic and manual segmentations are spatially similar to the degree expected between two qualified human neuroradiologists.
Lesion detection (F1 score): Instance-level F1 scores were 76% (95% CI: 69-83%) internally and 67% (95% CI: 59-74%) externally. The external F1 drop reflects the higher average lesion count per scan in the external cohort (5.42 vs. 3.68), particularly the greater prevalence of small, poorly defined infracentimetric lesions at some external centers. Among the seven external centers, performance varied: center VII achieved the highest Dice score of 0.92 (95% CI: 0.89-0.94), while center III had the lowest at 0.75 (95% CI: 0.55-0.89), largely attributable to a higher proportion of difficult lesion morphologies in that cohort.
Comparison to prior work: The only prior study specifically evaluating automated PCNSL segmentation (Pennig et al., 2021) used a model trained on glioblastoma data (DeepMedic) applied to 43 PCNSL patients and achieved a median Dice score of 0.76 without external validation. The current model outperforms that benchmark and adds the critical dimension of external multicenter validation, which is prerequisite for any real-world clinical tool. Zhang et al. (2021) included 92 lymphoma patients in a brain tumor classification pipeline but did not evaluate segmentation accuracy directly.
A critical clinical application of automated segmentation is the quantification of total tumor volume, which underpins treatment response assessment and is increasingly used as a prognostic biomarker. The study evaluated volumetric agreement between automatic and manual segmentations using Pearson correlation and Bland-Altman analysis. Pearson correlation coefficients were 0.99 (P less than .001) on the internal test set and 0.98 (P less than .001) on the external test set, indicating near-perfect linear agreement across the full range of tumor volumes encountered in clinical practice.
Robustness at high volumes: After excluding cases with very large tumor volumes (30 cm3 or greater), correlation coefficients remained high at 0.98 (internal) and 0.94 (external), confirming that the strong volumetric correlation is not driven solely by the largest, easiest-to-segment tumors. Bland-Altman analysis confirmed good overall agreement, though it revealed a slight systematic underestimation of tumor volume by the automatic model relative to manual segmentation. This underestimation is clinically important context: while the model provides highly correlated volume estimates, absolute volume values from automated segmentation should be interpreted with awareness of this systematic tendency, particularly when used to monitor fine-grained volume changes over treatment cycles.
Why volumetric assessment matters: Manual tumor volume measurement in PCNSL is time-consuming and prone to interobserver variability, which is a recognized challenge even in central review settings for clinical trials. A recent study (Lauer et al., 2024) found that MRI-based 3D tumor volume assessment improved early outcome prediction in CNS lymphoma patients, reinforcing the prognostic value of accurate volumetric data. The current model's ability to deliver near-manual-quality volume measurements automatically could meaningfully reduce measurement variance in both clinical practice and multi-institutional research settings.
The model also has potential to support radiomic analyses, since automated segmentation enables extraction of quantitative imaging features (shape, texture, intensity distribution) from the segmented tumor volume without the variability introduced by differing manual annotation styles across institutions. Integrating radiomic features with clinical and molecular data represents a logical next step for this segmentation tool.
One of the most clinically informative aspects of this study is the stratified performance analysis across five PCNSL imaging pattern categories. The model performed best on homogeneous and discrete enhancing lesions: group A (unique homogeneous lesion, Dice 0.84 internal / 0.93 external) and group B (multiple discrete homogeneous lesions, Dice 0.83 internal / 0.87 external). These patterns represent the most common PCNSL presentation, accounting for the majority of cases in both test sets, and also the pattern for which contrast enhancement produces the clearest tumor-background boundary on T1-weighted MRI.
Heterogeneous and perivascular patterns: Performance declined for group D (multiple heterogeneous lesions with linear ependymal enhancement) and most markedly for group E (linear perivascular enhancement), with group E showing a Dice score of 0.93 internally (only 1 case in the internal test set, limiting statistical interpretation) but substantially worse F1 scores. Perivascular and leptomeningeal enhancement patterns are inherently difficult to delineate because the enhancing structures are thin, linear, and poorly defined, without a clear focal mass. The authors note that boundary definition for these patterns is also challenging for experienced clinicians, and interleukin-10 quantification in cerebrospinal fluid is sometimes relied upon in complex imaging scenarios to support diagnosis.
Failure mode analysis: Cases with Dice scores below 0.60 were systematically reviewed. The most common error sources were: strong motion artifacts degrading image quality, low lesion contrast against background brain parenchyma, and undersegmentation of small lesions smaller than 1 cm, particularly those in periventricular locations where partial-volume effects and proximity to enhancing choroid plexus can reduce specificity. These failure modes align with known challenges in radiologist interpretation and underscore that the model reflects, rather than transcends, current imaging limitations.
Posttreatment cases: A subgroup analysis excluding posttreatment or unknown-treatment-status cases (5 internal, 14 external) showed no significant change in performance: Dice scores remained 0.84 (internal) and 0.87 (external), with no statistically significant difference from the full-dataset results (P = .81 and P = .82, respectively). This is clinically reassuring, though the small number of posttreatment cases limits firm conclusions about the model's suitability for monitoring patients under therapy.
Sample size constrained by disease rarity: Despite being the largest PCNSL segmentation dataset assembled to date, the cohort of 135 patients is small by deep learning standards. The rarity of PCNSL inherently limits training data availability, and the authors acknowledge this may restrict the model's ability to learn from uncommon imaging patterns, particularly very rare presentations such as purely meningeal spread. No patients with isolated leptomeningeal disease were included, meaning the model was neither trained nor validated for this pattern and should not be applied to it without further study. The open-source release of the model is intended to invite community contributions that could expand the training set over time.
T1-only framework: The segmentation model uses only postcontrast T1-weighted MRI, consistent with IPCG criteria that define PCNSL tumor burden by enhancing lesion volume. However, brain lymphoma also infiltrates tissue beyond the blood-brain barrier disruption visible on T1-weighted sequences, with additional tumor burden visible as hyperintense regions on T2-weighted FLAIR images. Because FLAIR hyperintensities in PCNSL are nonspecific (particularly in older patients who may also have vascular leukoencephalopathy or methotrexate-related leukoencephalopathy), the authors chose not to include them in the current framework. Expanding the model to incorporate FLAIR sequences in future work could provide a more complete picture of total tumor burden, particularly for treatment monitoring.
Posttreatment generalizability: Although the subgroup analysis suggests stable Dice scores excluding posttreatment cases, the relatively small number of posttreatment MRI examinations in the dataset (10 posttreatment cases internally, 13 externally) limits the ability to rigorously assess model performance in the treatment monitoring setting. Posttreatment PCNSL can present with residual enhancement, pseudo-progression, or therapy-related signal changes that differ morphologically from baseline disease, and model behavior in these situations warrants dedicated evaluation in expanded datasets.
Single-annotator reference standard: All segmentations were performed by a single primary annotator, with interreader variability assessed on only 50 of 135 cases. While the reported interreader metrics (Dice 0.81, MASD 1.64 mm) demonstrate reasonable agreement, a multi-reader consensus annotation protocol for the full dataset would provide a more robust reference standard. Additionally, the study did not conduct a formal power analysis to determine the minimum sample size needed for reliable model training, a methodological gap common in medical imaging deep learning studies.
The authors released the trained model as an open-source tool on GitHub (https://github.com/GuanghuiFU/nnunet_lymphoma_segment/tree/main), making it immediately usable by clinical researchers worldwide. Because the model uses the standard nnU-Net V2 framework without modifications, it is directly compatible with existing nnU-Net infrastructure, requiring no specialized deployment expertise. This positions the tool for rapid adoption in clinical research settings and for iterative refinement as additional PCNSL datasets become available from other institutions and national networks.
Multimodal expansion: The most immediate technical roadmap involves adding T2-weighted FLAIR sequences to the segmentation framework, enabling quantification of nonenhancing tumor burden alongside the current enhancing volume. This multimodal approach would align with emerging evidence that a more complete picture of PCNSL infiltration, beyond blood-brain barrier disruption alone, may improve response assessment sensitivity. Standardized volumetric measurements from a multimodal model could reduce the interobserver variability that currently complicates central review in PCNSL clinical trials.
Radiomic and molecular integration: Beyond tumor burden quantification, the segmentation masks produced by this model could serve as the foundation for radiomic feature extraction, linking quantitative imaging texture and shape features to clinical outcomes, molecular subtypes, or treatment response. Recent PCNSL research has identified four molecular subtypes with distinct clinical behaviors; a model that integrates automated radiomic features with molecular profiling and clinical data could provide more granular prognostic stratification than current clinical scoring systems. Survival prediction, therapeutic resistance modeling, and patient stratification for clinical trials are all plausible downstream applications.
Longitudinal monitoring and clinical integration: The ultimate clinical vision is a fully automated PCNSL MRI evaluation pipeline in which sequential scans are segmented, volumetric changes are tracked, and response to treatment is objectively assessed without manual radiologist delineation at every time point. This could reduce the measurement variability that currently hampers both routine response assessment and the interpretation of imaging endpoints in multicenter trials. Prospective validation of the model embedded in clinical workflows, with endpoints measuring agreement with expert-adjudicated response assessments, will be required before regulatory submission and routine clinical deployment.