Glioblastoma (GBM), primary central nervous system lymphoma (PCNSL), and solitary brain metastasis (BM) are among the most common intracranial neoplasms in adults, accounting for approximately 14.6%, 1.9%, and 17% of all primary brain and central nervous system tumors, respectively. Despite their different cells of origin, molecular biology, and required treatments, all three can produce overlapping appearances on standard MRI scans, creating a clinically consequential diagnostic dilemma for neuroradiologists and neurosurgeons.
Treatment divergence: The stakes of misdiagnosis are high because treatment pathways differ substantially across these three entities. Current guidelines recommend maximal safe surgical resection combined with concurrent chemoradiation therapy for GBM, maximal surgical resection followed by systemic therapy or stereotactic radiosurgery for BMs, and methotrexate-based chemotherapy combined with whole-brain radiotherapy for PCNSLs. Performing maximal surgical resection in a PCNSL patient - the correct intervention for GBM or BM - is both unnecessary and potentially harmful, as PCNSL responds to chemotherapy without requiring debulking.
MRI overlap and atypical presentations: On T1-weighted gadolinium-enhanced (T1Gd) images, GBMs typically show peripheral rim enhancement with central necrosis, a pattern that can closely resemble solitary BMs. PCNSLs usually exhibit homogeneous enhancement, but atypical PCNSLs with central necrotic cores can mimic GBMs. Advanced MRI modalities such as perfusion imaging, which detects the characteristically reduced cerebral blood volume (CBV) in PCNSL versus the elevated CBV in GBM, can help in typical cases but introduce operator-dependence, additional equipment requirements, and interpretive expertise barriers that limit worldwide applicability.
Stereotactic biopsy remains the diagnostic gold standard but carries a complication rate of up to 13%. Furthermore, preoperative steroid use, which is standard to reduce cerebral edema in GBM and BM patients, can suppress lymphoma histology in PCNSL tissue samples, yielding false-negative biopsies. This pilot study from the Fondazione IRCCS Ca Granda Ospedale Maggiore Policlinico in Milan evaluated whether a deep learning algorithm trained exclusively on routine T1Gd MRI could reliably differentiate these three tumor types and provide actionable surgical decision support.
The research team retrospectively reviewed records from 254 consecutive patients with histologically confirmed GBM, PCNSL with atypical radiological features, or BM who underwent preoperative brain MRI between June 2015 and April 2021. The deliberate focus on atypical PCNSL cases - specifically immunocompetent patients presenting with a central necrotic core, atypical anatomical location, or increased relative CBV - was a key methodological decision: these are precisely the cases where conventional radiology fails most frequently and where a decision-support tool would provide the greatest clinical benefit.
Exclusion criteria and final cohort: Patients were excluded for absent or inadequate MR images, prior intracranial intervention (surgical, gamma knife, or radiation therapy) that could alter imaging characteristics, and the presence of multiple enhancing lesions (which would indicate metastatic spread rather than a single indeterminate mass). AIDS-related and Epstein-Barr virus-related PCNSL subtypes were also excluded to reduce cohort heterogeneity. After applying these criteria, 121 patients were selected: 47 glioblastomas (age: 61.3 years IQR 48.9-73.7), 37 PCNSLs (age: 51.1 years IQR 43.3-58.9), and 37 BMs (age: 59.5 years IQR 49.9-69.1). The male-to-female ratio was 50:71 (58.3% female). Statistically significant differences in tumor volume were observed across groups (p=0.03), with median volumes of 56.31 cc for GBM, 39.00 cc for PCNSL, and 56.50 cc for BM.
BM primary sites: The 37 brain metastasis cases represented a clinically realistic mix of primary tumor subtypes: 14 (37.8%) lung cancers, 9 (24.3%) breast cancers, 6 (16.2%) colorectal cancers, 5 (13.5%) melanomas, and 3 (8.1%) endometrial cancers. This distribution mirrors real-world epidemiology of brain metastases and ensures the model was exposed to the radiological heterogeneity that characterizes BMs from different primary tumors.
Training and test split: The 121-patient cohort was divided following a 70/30 ratio into a training set used for deep learning model development and a hold-out test set used exclusively for internal validation. Data splitting was performed per patient rather than per image slice, a critical design choice that prevents data leakage - where slices from the same patient appear in both training and test sets, artificially inflating performance metrics. In total, 3,597 axial slices and regions of interest (ROIs) were extracted: 1,481 from GBM patients, 1,073 from PCNSL patients, and 1,043 from BM patients.
All MRI studies were acquired on a single 3 Tesla Philips Achieva scanner using a standardized 32-channel head coil protocol, which included axial T2-weighted sequences, 3D FLAIR, diffusion-weighted imaging (DWI) with b-values of 0-1000 sec/mm2, and gadolinium-enhanced (Gadovist 0.1 mL/kg or Prohance 0.2 mL/kg) axial and 3D T1-weighted sequences. The model was trained and validated exclusively on contrast-enhanced T1-weighted (CE-T1) sequences rather than multiparametric MRI, reflecting the practical reality that CE-T1 is the most universally available protocol worldwide.
Segmentation and preprocessing: Trained personnel manually delineated tumor regions of interest (ROIs) on every axial CE-T1 slice where the mass was visible using the Horos DICOM Viewer (version 3.3.5, open-source). Manual segmentation was selected over automated tools because current automated segmentation algorithms designed for one tumor type are not directly applicable to a multi-class scenario without introducing differential bias. After volume acquisition and signal intensity normalization, each ROI was centered in a 224x224-pixel black background frame and exported as PNG files for model input - a standardized format that removes surrounding anatomy and forces the network to focus exclusively on the lesion's internal characteristics.
ResNet-101 architecture and training: The deep neural network (DNN) was built on ResNet-101, a 2D convolutional neural network with 101 layers organized into 3-layer residual blocks, pre-trained on the ImageNet database. Transfer learning was applied by freezing the convolutional and pooling layers to preserve ImageNet-derived feature extraction capacity, while fine-tuning only the fully connected classification layer with the tumor MRI training data. Hyperparameters were set as follows: batch size 32, dropout rate 0.25, rectifier linear unit (ReLU) activation function, 50 training epochs, stochastic gradient descent with Adam optimizer, initial learning rate 0.005, and batch normalization at each layer to improve training stability.
Majority voting and Grad-CAM: Because each patient contributed multiple image slices, the final diagnostic class was determined by majority vote across all ROIs from that patient - whichever class received the most votes across all slices was assigned as the patient-level prediction. To address the "black box" concern inherent in deep learning, the team implemented Gradient-weighted Class Activation Mapping (Grad-CAM) visualization, which highlights the spatial regions within each image that contributed most to the model's prediction, providing a degree of interpretability that can be inspected by clinicians.
The trained ResNet-101 DNN was evaluated on the hold-out test set using a One-vs-Rest (OvR) multiclass strategy, which generates binary classification metrics for each tumor class independently against the combined other two. Performance is reported as average values with 95% bootstrap confidence intervals. The model achieved its strongest performance for atypical PCNSL, followed by GBM, with BM classification presenting the greatest difficulty - a pattern that has meaningful clinical implications given the specific diagnostic challenges each entity presents.
PCNSL classification: The model demonstrated near-optimal discrimination for PCNSL with an AUC of 0.98 (95% CI: 0.95-1.00), accuracy of 94.65% (95% CI: 89.19%-100.00%), precision (PPV) of 91.57% (95% CI: 76.92%-100.00%), NPV of 93.54% (95% CI: 87.38%-100.00%), sensitivity of 91.03%, specificity of 96.23%, and F1-score of 0.91. The high NPV of 93.54% means the model is highly reliable at ruling out PCNSL when it predicts a different class, which is particularly relevant for clinical decision-making because a negative PCNSL prediction meaningfully shifts the differential toward a surgically resectable lesion.
Glioblastoma classification: GBM differentiation was classified at AUC 0.90 (95% CI: 0.81-0.97), accuracy 83.08% (95% CI: 72.83%-91.89%), PPV 75.50%, NPV 79.92%, sensitivity 80.01%, and specificity 81.84%, with F1-score 0.77. Performance was moderate to good across all metrics, reflecting the fact that GBM occupies a challenging intermediate position between the more morphologically distinctive PCNSL and the heterogeneous BM cohort.
Brain metastasis classification: BM was the most difficult class, with AUC 0.81 (95% CI: 0.70-0.95), accuracy 81.07%, PPV 71.11%, and sensitivity 63.61%. However, the model achieved high specificity of 88.46% and NPV of 84.37% for BM, meaning it was substantially more reliable at correctly excluding BM than at confirming its presence. The authors attribute the relatively lower sensitivity to the inherent histological heterogeneity of brain metastases arising from multiple different primary tumor types, each potentially producing distinct MRI signatures within the single BM training category.
Beyond three-class tumor differentiation, the authors conducted a secondary grouped binary analysis that has direct clinical utility: distinguishing surgically resectable lesions (GBM or BM) from non-resectable ones (PCNSL). This framing reflects the actual neurosurgical decision point, where the critical question is not always which specific tumor type is present, but rather whether the patient should proceed to maximal surgical resection or diagnostic biopsy. PCNSL, which responds to chemotherapy and does not benefit from debulking, represents the "non-resectable" category, while both GBM and BM benefit from maximal safe resection.
Binary resectability performance: The DNN achieved outstanding performance on this clinically structured binary task: AUC 0.92 (95% CI: 0.83-0.99), accuracy 94.72% (95% CI: 89.19%-100.0%), PPV 91.88% (95% CI: 78.57%-100.0%), NPV 94.76% (95% CI: 86.13%-100.00%), sensitivity 90.84%, specificity 96.34%, and F1-score 0.91. These metrics indicate that the model could correctly direct approximately 95 out of 100 patients toward the appropriate surgical strategy - either proceeding to resection or biopsy - based solely on the T1Gd MRI appearance.
The near-perfect specificity of 96.34% for the resectability analysis is particularly noteworthy. A high specificity means very few PCNSL patients would be incorrectly labeled as having a resectable tumor and therefore unnecessarily subjected to major brain surgery. The NPV of 94.76% means that when the model predicts a lesion is non-resectable (PCNSL), there is a 94.76% probability that this prediction is correct, providing clinicians with meaningful confidence to defer aggressive surgery.
Computation speed: A practical advantage highlighted by the authors is the model's computation speed of under 10 seconds per patient, which makes real-time intraoperative or preoperative decision support feasible. This is in sharp contrast to advanced MRI acquisition and analysis workflows, which require specialized hardware, extended acquisition time, and expert postprocessing. The low computational overhead also means the model could run on modest institutional hardware without requiring large computing clusters.
To contextualize the DNN's performance, the team retrospectively computed the diagnostic performance of senior neuroradiologists with at least 10 years of dedicated experience on the same patient cohort. Critically, the radiologists had access to the complete multisequence DICOM package including T1-weighted, T2-weighted, FLAIR, diffusion-weighted, CE-T1, dynamic contrast-enhanced, and perfusion sequences - a substantially richer data source than the single T1Gd sequence available to the model. Each radiologist independently classified every tumor, and their retrospective results from clinical report charts were used as the gold standard for comparison.
Radiologist performance by tumor type: The neuroradiologists achieved overall accuracy of 84.38% for PCNSL (sensitivity 55.56%, specificity 91.03%, PPV 58.82%, NPV 89.87%), 85.87% for GBM (sensitivity 82.35%, specificity 80.43%, PPV 93.33%, NPV 80.43%), and 91.67% for BM (sensitivity 85.71%, specificity 93.33%, PPV 78.26%, NPV 95.89%). The most striking finding was the low PCNSL sensitivity of 55.56% for radiologists - more than one in three atypical PCNSLs was misclassified, likely as GBM or BM, leading to subsequent inappropriate surgical planning.
Head-to-head comparison: The DL model demonstrated an accuracy improvement of approximately +14% for PCNSL (94.65% vs. 84.38%) and approximately +5% for GBM (83.08% vs. 85.87%, noting the model was slightly below radiologists here) compared to the gold standard. The model underperformed radiologists by approximately 10% for BM (81.07% vs. 91.67%). The authors note this BM gap is unsurprising: experienced radiologists integrate clinical history, known primary cancer diagnoses, and the multiparametric MRI findings that the model was not given access to. For BM, context is often decisive.
Complementary strengths: The comparison reveals a clinically important pattern: the DNN excels precisely where expert radiologists struggle most - atypical PCNSL cases that mimic GBM - while radiologists hold an advantage where clinical context matters most (BM). This suggests a complementary rather than competitive role for the AI tool, which could serve as a specific alert system for atypical PCNSL cases that are at high risk of misclassification by conventional radiological assessment.
Sample size and heterogeneity coverage: The cohort of 121 patients, while larger than many comparable studies in the literature, remains relatively limited for capturing the full spectrum of radiological variability that GBMs, PCNSLs, and BMs exhibit in real clinical settings. The deliberate inclusion of only atypical PCNSL cases - a design choice intended to maximize clinical relevance - also means the model was not exposed to the full range of typical PCNSL presentations, potentially limiting generalizability to mixed clinical populations where both typical and atypical cases are encountered.
Monocentric design and scanner dependence: All imaging was acquired on a single 3T Philips Achieva scanner at one institution using a standardized protocol. Deep learning models are known to be sensitive to differences in scanner hardware, vendor, field strength, and acquisition parameters. A distribution shift in input data - for instance, when images are acquired on a different MRI platform at a different center - can alter the statistical patterns in pixel intensities that the network relies upon, potentially degrading performance. The authors acknowledge that external validation at institutions using different scanners is a prerequisite before any clinical deployment.
Manual segmentation as the preprocessing bottleneck: The pipeline requires trained personnel to manually delineate tumor ROIs on every axial CE-T1 slice before model inference. This is time-consuming, subject to interobserver variability, and represents a practical barrier to rapid clinical deployment. The authors considered automated segmentation but noted that available automated tools are typically optimized for a single tumor type and would introduce differential biases across the three classes in a multi-tumor scenario. Manual segmentation was therefore retained as the current gold standard, but it limits the scalability of the approach.
Black-box interpretability: Despite the implementation of Grad-CAM visualization to highlight which image regions drive predictions, deep neural networks fundamentally operate through learned statistical representations that are not directly interpretable in terms of the morphological criteria pathologists and radiologists use clinically. Grad-CAM provides a heatmap overlay but does not fully explain why specific pixel patterns trigger one classification over another, limiting clinician confidence and making it difficult to identify and correct systematic errors in model behavior.
Positioning against prior machine learning studies: Before this study, the published literature contained machine learning models that addressed binary differentiation (GBM vs. PCNSL, or GBM vs. BM) but no three-class model simultaneously distinguishing all three entities. Kunimatsu et al. reported a support vector machine using radiomic features that achieved 0.75 accuracy for GBM vs. PCNSL. Xia et al. designed a CNN on multi-parametric MRI sequences achieving 0.884 accuracy for the same binary problem. McAvoy et al. used an EfficientNetB4 model on whole-brain scans without segmentation, achieving comparable accuracy (GBM: 0.94, PCNSL: 0.95), but the whole-brain approach without segmentation risks the network learning spurious features tied to anatomical location and laterality in unbalanced training datasets rather than true lesion biology.
Why segmentation matters: The ROI-based approach used in this study forces the model to learn from the lesion's intrinsic imaging characteristics rather than its anatomical context. PCNSLs preferentially occur in periventricular and deep white matter locations, and a model trained on unsegmented whole-brain images could achieve high accuracy by learning anatomical location as a proxy for tumor type - a shortcut that would fail catastrophically for atypical cases occurring in unusual locations, which are precisely the cases this study was designed to address.
Clinical use case and workflow integration: The authors envision the model as a decision-support tool specifically for atypical cases - not as a replacement for neuroradiological assessment, but as an adjunct that flags ambiguous lesions and recommends either surgical resection or biopsy. With a computation time under 10 seconds per patient, the model is fast enough to be embedded into preoperative planning workflows. The training on open-source Python packages (PyTorch, Numpy, Scikit-learn) and a commodity GPU (Tesla K80 12GB) means the infrastructure requirements are accessible to institutions without high-performance computing clusters.
Future directions: The authors call for prospective multicenter validation using imaging from multiple scanner platforms and institutions to test generalizability. They also propose a prospective randomized study comparing neurosurgical and neuroradiological decision-making with and without AI tool access, measuring concrete clinical outcomes including diagnostic accuracy, time to definitive diagnosis, inappropriate surgical procedures avoided, and patient outcomes. Integration of additional MRI sequences (DWI, perfusion, FLAIR) into a multiparametric model is identified as a likely pathway to close the performance gap for BM classification, where clinical context and multi-sequence analysis currently give experienced radiologists a decisive advantage.