Primary central nervous system lymphoma (PCNSL) is an exceedingly rare intracranial tumor, accounting for only 4 to 7% of newly diagnosed intracranial lesions. It is B-cell lymphoma confined to the brain, spinal cord, leptomeninges, or eyes, and it carries a markedly different treatment trajectory than glioblastoma (GBM), the most common malignant intraaxial tumor responsible for nearly half of all newly diagnosed intracranial malignancies. Although the two conditions demand entirely different therapeutic approaches, they can be remarkably difficult to distinguish on preoperative MRI alone, even for experienced neuroradiologists and neurosurgeons.
Shared MRI features: On contrast-enhanced MRI, PCNSL and GBM share a range of morphological characteristics. Both can present with large perifocal edema, inhomogeneous contrast enhancement, and in some cases central necrosis. These overlapping features make confident preoperative classification unreliable from imaging alone, and misclassification carries real clinical consequences. When PCNSL is misidentified as GBM preoperatively, patients may receive corticosteroids such as dexamethasone for symptom relief, which, while effective at reducing cerebral edema, can lyse lymphoma cells and render subsequent stereotactic biopsy non-diagnostic, permanently obscuring the pathological diagnosis.
The clinical stakes: Stereotactic biopsy remains the gold standard for histopathological confirmation of intracranial lesions but is a specialized procedure available only at large centers. Clinicians at peripheral hospitals therefore face a dilemma: treat patients with corticosteroids for symptom relief (risking an inconclusive biopsy if the lesion is PCNSL) or withhold treatment until a biopsy can be performed (forcing patients to endure debilitating symptoms of cerebral edema). A reliable preoperative imaging-based classifier could resolve this dilemma by flagging high-probability PCNSL cases where corticosteroid use should be deferred until tissue diagnosis is secured.
This 2024 study from Heidelberg University Hospital, published in iScience, aimed to develop an automated deep-learning diagnostic workflow based on a convolutional neural network (CNN) to preoperatively distinguish PCNSL from GBM using standard clinical MRI sequences, and to rigorously benchmark its performance against trained human experts.
The study retrospectively reviewed patient records at Heidelberg University Hospital spanning January 2008 to August 2023, identifying patients who had undergone stereotactic biopsy with pathologically confirmed diagnoses. A total of 158 patients were enrolled: 76 male and 82 female, mean age 65.6 plus or minus 12.5 years. From this cohort, 68 PCNSL and 69 GBM patients formed the training set, and 20 patients (10 PCNSL, 10 GBM) were randomly selected as a held-out external test set. Critically, the test set patients were treated as unknown to the model throughout training, providing a true external validation rather than merely internal cross-validation.
MRI inclusion criteria: Only patients with complete, high-resolution MRI datasets acquired within one week before stereotactic surgery were included. All four standard clinical sequences were required: T1, contrast-enhanced T1 (T1-CE), T2, and fluid-attenuated inversion recovery (FLAIR). The decision to restrict to these four widely available sequences was deliberate, as the authors sought a workflow that could be deployed at any center without requiring specialized acquisition protocols such as diffusion, perfusion, or spectroscopy sequences.
Preprocessing pipeline: All images underwent automated skull-stripping using FSL BET (for T1 images) and the HD-BET neural network algorithm (for remaining modalities). Images were then registered to the MNI152 standard brain template using FMRIBS's Linear Image Registration Tool (FLIRT), with the FLAIR sequence used as the registration reference. This standardized spatial normalization ensured that all images were comparable across the dataset regardless of scanner hardware or acquisition parameters.
Automated tumor segmentation: A nnU-Net model trained on the BraTS2021 benchmark dataset was used to generate automated segmentation masks for three tumor compartments: peritumoral edema (ED), non-enhancing tumor (NET), and enhancing tumor (ET). All segmentation masks were manually reviewed and vetted by clinicians before being passed to the classification network. Volumetric analysis confirmed no statistically significant differences in tumor volume, edema volume, or enhancing tumor volume between PCNSL and GBM groups, or between training and test subsets, validating the comparability of the cohorts.
The automated classification algorithm was built using the Medical Open Network for Artificial Intelligence (MONAI) framework with PyTorch as the backend. Two families of convolutional neural network architectures were systematically evaluated: ResNet (Residual Networks) and DenseNet (Densely Connected Convolutional Networks). ResNet architectures operate by learning residual mappings through skip connections that add the input directly to the output of a block, helping mitigate the vanishing gradient problem in deep networks. DenseNet extends this concept further by connecting each layer to every subsequent layer in a feed-forward fashion, maximizing feature reuse across the network depth.
Hyperparameter search: Five ResNet depths were evaluated: ResNet10, ResNet18, ResNet34, ResNet50, and ResNet101. The two best-performing ResNet variants (ResNet10 and ResNet18) were taken forward for hyperparameter tuning. Similarly, DenseNet169 and DenseNet264 were subjected to hyperparameter optimization. All models were trained with a learning rate of 0.0001 over 150 epochs using 5-fold internal cross-validation with ensemble prediction. Monte Carlo Dropout was implemented to assess model uncertainty, and Binary Cross-Entropy with Logits loss (BCEL) was used as the optimization objective.
Architecture comparison results: Across the grid search cross-validation, DenseNet169 achieved the best performance score of 0.6368, compared to 0.6242 for both ResNet10 and ResNet18, and 0.6220 for DenseNet264. DenseNet169 was selected as the final model and subjected to full hyperparameter optimization before evaluation on the held-out test set. The superior performance of DenseNet169 over shallower ResNet variants suggests that the denser connectivity pattern provides informative features for distinguishing the subtle MRI morphological differences between PCNSL and GBM.
Saliency map analysis: To interrogate where the network focused its attention, saliency maps were generated using two complementary interpretability techniques: integrated gradients and guided backpropagation. Visual inspection of the maps showed the network consistently directing attention toward the tumor region even when segmentation masks were not provided. Statistical comparison of saliency-to-tumor-overlap between correctly and incorrectly classified cases showed no significant difference (p = 0.345, unpaired t-test), suggesting the algorithm attended to relevant anatomical regions even in misclassified cases.
Six experienced physicians from Heidelberg University Hospital served as human raters: neurosurgeons and neuroradiologists blinded to the histopathological diagnosis. Each rater independently reviewed the MRI data for all 20 test cases and all 120 cases in the training set (20 cases times 6 raters), completing a 12-item questionnaire that captured their diagnosis, their confidence level on a percentage scale, and their assessment of the relative importance of each imaging modality and lesion characteristic. Raters first reviewed only T1, T1-CE, T2, and FLAIR sequences, then were given access to all available imaging sequences and could revise their diagnosis.
Overall accuracy: When considering all 120 predictions across 6 raters on 20 test cases, the overall human accuracy was 79.16% (95 out of 120 correct). Accuracy on PCNSL cases (49 out of 60 correct) was marginally higher than on GBM cases (46 out of 60), though this difference did not reach statistical significance (p = 0.5, chi-square test). The ROC-AUC across all human predictions was 0.8091, with a higher AUC for PCNSL detection (0.9) than for GBM detection (0.603), reflecting a relative asymmetry in diagnostic confidence.
Experience does not predict accuracy: A perhaps surprising finding was that years of experience with cranial MRI did not significantly correlate with accuracy for either GBM (F = 0.59, p = 0.48) or PCNSL (F = 0.08, p = 0.79) diagnosis. Similarly, no significant difference in diagnostic accuracy was detected between radiologists and neurosurgeons. The mean time to prediction was 110.3 plus or minus 53 seconds and was not significantly different between correct and incorrect diagnoses (106.4 vs. 124.8 seconds, p = 0.5) or between PCNSL and GBM cases (p = 0.93).
Importance ratings: Raters consistently rated T1-CE (contrast-enhanced T1) as the most important MRI sequence for diagnosis, followed by FLAIR and T2. Native T1 was considered least important. Among tumor characteristics, contrast-enhancing components were ranked most relevant by all raters for both tumor types. Necrosis was rated significantly more important for GBM than for PCNSL (p = 0.017), reflecting a known morphological differentiator. When given access to additional sequences beyond the four standard ones, accuracy improved slightly to 86.67%, though some raters disimproved by overcorrecting initially accurate diagnoses.
After training on 137 cases (68 PCNSL, 69 GBM) using 5-fold cross-validation and ensemble prediction, DenseNet169 was applied to the 20 held-out test patients at a decision threshold of 0.5. The network correctly classified 16 out of 20 cases (80% accuracy), with two PCNSL cases and two GBM cases each misdiagnosed. The positive predictive value (PPV) and negative predictive value (NPV) for PCNSL were both 0.8. The ROC-AUC for automated classification on the test set was 0.9, representing a strong discriminative performance considerably above the 0.8091 AUC achieved by human raters across all their predictions.
Speed advantage: The mean time to prediction for the CNN was 0.045 plus or minus 0.0003 seconds, approximately 2,000 times faster than the mean human prediction time of approximately 110 seconds. This difference was not statistically different between correct and incorrect predictions (p values not significant) or between PCNSL and GBM cases, demonstrating consistent processing speed regardless of case difficulty.
The role of individual MRI modalities: To determine how much each imaging modality contributed to network performance, two ablation experiments were conducted. First, the network was trained and tested on only one modality at a time. Performance dropped dramatically under this condition, with T1 alone yielding the best single-modality AUC of 0.65, compared to 0.9 with all four modalities combined. Interestingly, T1-CE alone performed worst at AUC 0.47 (near chance), contrasting sharply with the human raters who rated T1-CE as their most informative sequence. This divergence suggests the CNN relies on different discriminative features than those consciously identified by clinicians.
Modality removal experiment: In a second ablation, each modality was removed one at a time while the remaining three were retained. Removing FLAIR caused the largest single-modality performance drop, with AUC declining from 0.9 to 0.71. Removing T2 had the second-largest effect. Removing T1-CE or native T1 had comparatively minor effects. This finding stands in contrast to the human raters' reliance on T1-CE and suggests the network leverages information present in T2 and FLAIR that may not be explicitly recognized as diagnostically informative by human experts.
The most striking finding of the study was that combining human and automated classification yielded substantially higher accuracy than either approach alone. In cases where human raters reached a consensus diagnosis (15 out of 20 cases), the consensus accuracy was 19 out of 20 (95% accuracy). For the remaining 5 cases where raters dissented (disagreed among themselves), the CNN was invoked as the tiebreaker. When the CNN's classification was used for these 5 dissent cases, the combined approach achieved an overall accuracy of 95% (19 out of 20 cases correct), with only a single misclassification.
Complementary error profiles: A particularly important observation was that the CNN and human raters made mistakes on different cases. The network correctly predicted cases that human raters struggled with and agreed were ambiguous (the dissent cases), while the human raters correctly diagnosed cases that the network misclassified (the unequivocal cases). This complementarity is not merely coincidental; it implies that the network is learning a fundamentally different set of diagnostic features than those used by clinicians, making the combination more robust than either component alone.
Clinical workflow implications: The layered analysis proposed here has a natural clinical translation. In practice, cases where a team of specialists reaches unanimous agreement can be classified with high confidence using the human diagnosis alone. Cases generating disagreement among reviewers, precisely those most prone to error under the current standard of care, can be flagged for CNN review. Rather than replacing the clinician, the algorithm acts as an objective adjudicator for diagnostically challenging ambiguous cases, the subset where the current 73 to 80% human accuracy leaves the most room for improvement.
This hybrid approach achieved an NPV for PCNSL of 91% and a sensitivity of 90% for PCNSL, as well as 100% sensitivity and NPV for GBM, when considering only the test subset. These numbers reflect the potential of a well-implemented human-AI collaboration to substantially reduce the rate of misclassified lymphoma cases prior to treatment decisions.
Artificial task framing: Although raters were blinded to individual patient diagnoses, they were aware that the study was specifically investigating the PCNSL vs. GBM distinction. This knowledge likely primed them to look for PCNSL-specific features they might not scrutinize as carefully in routine clinical practice and to restrict their differential to just these two entities. In everyday neurosurgical and neuroradiological practice, the differential diagnosis for an intracranial lesion is considerably broader, and clinicians must actively entertain metastases, high-grade astrocytoma, abscess, and other entities. The binary framing likely inflated human performance relative to real-world accuracy.
Retrospective design and missing sequences: Because the study is retrospective, many PCNSL patients who underwent biopsy could not be included because of missing MRI sequences from the required four. This selection bias means the enrolled cohort may not be representative of all PCNSL patients seen at the institution. Furthermore, the study deliberately excluded additional sequences known to be informative for PCNSL/GBM distinction, particularly diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) maps, because these were not routinely available for referred patients with preexistent imaging. Including these sequences in future work could improve CNN performance further.
GBM case selection: The study enrolled GBM patients who had undergone stereotactic biopsy, a procedure typically reserved for diagnostically uncertain cases. GBMs with unequivocal imaging characteristics typically proceed directly to resection without biopsy. Consequently, the GBM cohort likely represents an enriched population of cases with atypical or ambiguous features that resemble other entities, making the classification task harder than if randomly selected GBMs were used. This conservative selection is appropriate for modeling real-world clinical need but means the reported accuracies may not be directly comparable to studies using unselected GBM populations.
Small external test set: The external validation cohort was limited to 20 patients (10 per class), approximately 13% of the training dataset size. While this ratio is comparable to proportions used in other published studies in the field, it limits the statistical power to precisely estimate model performance. Confidence intervals around the 80% accuracy figure are correspondingly wide. Prospective validation on a larger multicenter cohort would be required before this tool could be considered for clinical deployment.
Expanding the case library: The most immediate need identified by the authors is prospective validation on a larger patient cohort. With PCNSL constituting only 4 to 7% of intracranial tumors, assembling sufficient training data requires multicenter collaboration over extended time periods. The authors note that pooling data and computational approaches across multiple neurosurgical centers would increase both training set size and the diversity of scanner hardware, patient populations, and acquisition protocols represented, directly improving model generalizability. The code and data underlying this study were deposited publicly on Figshare to facilitate future collaborative work.
Incorporating additional MRI sequences: The study's decision to restrict to four standard sequences was pragmatic, but future iterations could incorporate DWI, ADC, susceptibility-weighted imaging (SWI), and cerebral blood volume (CBV) sequences where available. Notably, when human raters in this study were given access to additional sequences beyond the four standard ones, they rated DWI and ADC as the most informative for PCNSL diagnosis. Incorporating these into the CNN may close the gap between the network's 80% accuracy and the 86.67% achieved by humans with access to supplementary sequences.
Broader tumor differentiation: The current binary PCNSL vs. GBM classifier addresses only one clinically important differential. Future work could extend the framework to distinguish among additional intracranial tumor entities, including brain metastases, high-grade astrocytomas, and demyelinating disease. Related published studies have begun exploring three-way classifiers (GBM, PCNSL, metastasis), and the current dataset and preprocessing pipeline could serve as a foundation for training such multi-class models.
Clinical integration pathway: The authors envision this tool as a decision support system accessible preoperatively to clinicians at institutions with or without subspecialty neuroradiology expertise. Because the preprocessing pipeline is fully automated and requires only standard MRI sequences, it could in principle be integrated into picture archiving and communication systems (PACS) or deployed as a web-based service that returns a diagnosis probability within seconds of image upload. Before clinical adoption, prospective trials evaluating the tool's impact on actual patient management decisions and outcomes will be essential.