Primary central nervous system lymphoma (PCNSL) and glioblastoma (GBM) are two of the most common malignant brain tumors in adults and the elderly, yet they represent biologically distinct entities with radically different treatment protocols and prognoses. PCNSL is a rare extranodal non-Hodgkin lymphoma confined to the brain, spinal cord, eyes, or leptomeninges, and is typically treated with high-dose methotrexate-based regimens and sometimes whole-brain radiation. GBM is the most aggressive primary brain tumor, typically treated with surgical resection followed by temozolomide chemotherapy and radiation. Confusing one for the other delays appropriate therapy, exposes patients to ineffective treatments, and can foreclose curative opportunities in PCNSL, where early chemotherapy without prior surgery can preserve neurological function.
The MRI overlap problem: Despite being distinct diseases, PCNSL and GBM have overlapping appearances on conventional MRI. Both typically present as contrast-enhancing lesions with surrounding edema, often in deep white matter or periventricular locations. PCNSL classically shows homogeneous enhancement, restricted diffusion, and a lower apparent diffusion coefficient (ADC) than GBM, but these features are not sufficiently discriminating when encountered in individual patients. Radiologists must rely on a combination of morphological features, patient demographics, and clinical context, a process that is subjective and experience-dependent.
The limitations of current practice: Special MRI sequences such as diffusion-weighted imaging (DWI), MR spectroscopy (MRS), and dynamic susceptibility contrast (DSC) perfusion imaging can provide additional discriminatory information, but these protocols are resource-intensive, not universally available, and subject to their own interpretive variability. As a result, many patients still require stereotactic brain biopsy for definitive diagnosis, an invasive procedure that carries risks including hemorrhage, neurological deficit, and, in PCNSL, potential diagnostic confusion if the patient has already been empirically treated with corticosteroids. This diagnostic dilemma, affecting thousands of patients annually, has motivated researchers to explore whether AI can achieve reliable non-invasive differentiation.
Tumor radiomics represents a quantitative imaging approach that extracts hundreds of mathematical features from segmented tumor regions, including measures of shape, intensity distribution, and texture, which are invisible to the naked eye. When fed into machine learning classifiers, these features can reveal patterns in imaging data that exceed human perceptual capacity. Deep learning approaches, particularly convolutional neural networks (CNNs), offer an alternative by learning discriminative features directly from raw image data without requiring manual feature engineering.
The authors followed PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines throughout the review process. An electronic search was conducted across Medline via PubMed, EMBASE, and the Cochrane Central Register of Controlled Trials and Database of Abstracts of Reviews of Effectiveness (DARE) from January 1990 through December 2021, with no language restrictions. The search strategy combined MeSH terms and keywords including "Magnetic Resonance Imaging," "primary central nervous system lymphoma," "brain lymphoma," "PCNSL," "glioblastoma," "diagnosis," "radiomics," "machine learning," "deep learning," and "artificial intelligence," linked using Boolean AND and OR operators. The search was supplemented by hand-searching review articles, cross-references, and conference proceedings to capture any studies not indexed in the major databases.
Eligibility criteria: Studies were included if they reported on ML or DL algorithms specifically designed to differentiate PCNSL from GBM using MR imaging. Studies comparing ML or DL performance against radiologists (rather than against histopathology as a gold standard) were deliberately excluded to maintain analytical homogeneity and to allow direct comparison of algorithm performance against the definitive reference standard. Editorials, commentaries, letters, case reports, and studies addressing PCNSL differentiation from other glioma grades (not specifically GBM) were also excluded. This focus on histopathological confirmation as the reference standard, rather than radiologist consensus, is an important methodological choice that limits bias from spectrum effects and ensures consistent ground truth across studies.
Data extraction and quality assessment: Two independent reviewers (AG and JSG) screened titles and abstracts, retrieved full texts, and extracted data using a standardized extraction sheet. A third reviewer (AM) independently verified extractions, with discrepancies resolved by consensus. Quality was assessed using two tools: QUADAS-2 (Quality Assessment of Diagnostic Accuracy Studies 2), a standard tool that evaluates four bias domains - patient selection, index test, reference standard, and flow and timing - rating each as low, high, or unclear risk; and the Radiomic Quality Score (RQS), a specialized quality tool for neuro-oncology radiomics studies that scores up to 36 points across six domains including image protocol quality, feature extraction, validation methodology, and open-science practices.
Statistical analysis: Meta-analysis was performed in R Studio version 3.6.1 using the 'meta' and 'mada' packages. Because heterogeneity was high (I-squared exceeding 50% for all outcomes), random-effects models were used throughout. The primary outcome metrics were pooled sensitivity, specificity, accuracy, and balanced accuracy (defined as the arithmetic mean of sensitivity and specificity). Hierarchical Summary Receiver Operating Characteristic (HSROC) curves were constructed using the mada package, consistent with Cochrane Collaboration recommendations for meta-analyses of diagnostic accuracy studies. Forest plots were generated for all four metrics. Threshold effect analysis used the Spearman correlation coefficient between sensitivity and false-positive rate.
From an initial pool of 70 identified studies, 38 were duplicate records. Of the remaining 32, 7 were excluded based on title and abstract review, leaving 25 full-text articles retrieved for detailed assessment. After applying inclusion and exclusion criteria, 10 studies were selected for the meta-analysis. These 10 studies enrolled a combined total of 1,311 patients (PCNSL and GBM combined), with individual study sample sizes ranging from 53 (Shrot et al., 2019) to 260 (Park et al., 2020). All 10 studies used retrospective designs; none had a prospective enrollment strategy. Sixty percent of studies originated from hospitals in Asia, particularly China and South Korea, with the remaining studies from Israel, Spain, and the United States.
Machine learning studies (6 studies): Six of the ten included studies used classical ML approaches, with support vector machines (SVM) appearing as the classifier of choice in at least half of them. Other classifiers used included logistic regression, random forest, naive Bayes, gradient-boosted decision trees (GBDT), XGBoost, linear discriminant analysis (LDA), and multilayer perceptron (MLP) neural networks. Radiomic features were typically extracted using established tools such as Pyradiomics or LifeX, from manually or semi-automatically segmented tumor regions of interest (ROIs). Feature selection methods varied widely, including distance correlation, LASSO (Least Absolute Shrinkage and Selection Operator), random forest feature importance, and minimum redundancy maximum relevance (mRMR). The Kim et al. (2018) study from South Korea was the only one that performed external validation on an independently collected dataset, using radiomics features from multi-parametric MRI (T1-weighted contrast, DWI, T2-weighted) with a logistic regression, SVM, and random forest ensemble.
Deep learning studies (4 studies): Four studies used convolutional neural networks (CNN) as their primary classification approach. Three of the four CNN studies used T1-weighted contrast-enhanced images as input, while one (Park et al., 2020) incorporated dynamic susceptibility contrast perfusion (DSC-PWI) time-signal intensity curves processed through an autoencoder architecture to capture perfusion heterogeneity patterns across the tumor volume. The CNN approach by Guoqing et al. (2018) used a patch-based sparse representation method applied to 3T MRI scans in a completely automated pipeline, without manual segmentation. Scale Invariant Feature Transformation (SIFT) features were used in one CNN study (Chen et al., 2018) as an intermediate high-throughput feature extraction step from conventional MRI sequences.
MRI protocols and scanner heterogeneity: Five studies used exclusively 3T MRI scanners, while three studies used both 3T and 1.5T scanners, and two studies did not report scanner specifications. The MRI sequences analyzed varied substantially: some studies used only T1-weighted contrast-enhanced images, while others incorporated DWI, T2/FLAIR, ADC maps, DSC perfusion, and diffusion tensor imaging (DTI). This variability in imaging protocols is a key source of heterogeneity across the meta-analysis. An imbalance in class representation was observed in most studies, with a PCNSL-to-GBM ratio of approximately 1:2 to 1:3 in favor of GBM, reflecting the relative rarity of PCNSL in clinical practice.
The diagnostic performance metrics across the ten included studies spanned a wide range, reflecting genuine methodological heterogeneity rather than a consistent signal. Accuracy ranged from 82.0% (Xiao et al., 2018, a radiomic ML study using naive Bayes, SVM, logistic regression, and random forest on 82 patients) to 97.9% (Chen et al., 2020, using five feature selection methods combined with LDA, SVM, and LR classifiers on 138 patients with T1-contrast radiomics via LifeX). Sensitivity - the ability to correctly identify PCNSL cases - ranged from a minimum of 78.0% (Xiao et al., 2018) to 100% (Shrot et al., 2019, a binary SVM study using perfusion DSC and DTI on 53 patients via leave-one-out cross-validation). Specificity - the ability to correctly identify GBM cases - ranged from 76.0% (Park et al., 2020, CNN with perfusion) to 100% (Shrot et al., 2019). Balanced accuracy ranged from 84.5% (Xiao et al.) to 100% (Shrot et al.).
AUC values: Area under the ROC curve was reported or calculable for 7 of the 10 studies and ranged from 0.89 to 0.982. The highest AUC of 0.982 was reported by Chen et al. (2018) using a CNN with SIFT features on 96 patients, while Xiao et al. (2018) reported the lowest AUC of 0.90 using a multi-classifier radiomic ML approach. The Kim et al. (2018) study, which is notable for being the only externally validated study, reported AUC of 0.956 with accuracy 94.7%, sensitivity 96.6%, and specificity 92.9%, making it one of the strongest and most credible results in the series. Park et al. (2020) could not report accuracy because of the study design, but achieved sensitivity 95.0% and specificity 76.0% with AUC 0.89.
Pooled meta-analysis results: The subgroup analysis using random-effects pooling across all 10 studies showed that the overall pooled sensitivity for identifying PCNSL was 92% (95% CI: 0.88-0.95) and pooled specificity was 94% (95% CI: 0.89-0.97). Importantly, neither sensitivity nor specificity showed statistically significant differences based on sample size (studies above versus below 100 patients). The HSROC curve, constructed to account for threshold effects across studies, showed a large divergence between the 95% confidence region and the 95% prediction region, visually illustrating the degree of between-study heterogeneity. The Spearman correlation coefficient between sensitivity and false-positive rate was -0.16 (p=0.66), confirming the absence of a threshold effect in the pooled data.
Heterogeneity assessment: Significant statistical heterogeneity was detected across all outcome measures: sensitivity I-squared = 83% (p less than 0.01), specificity I-squared = 87% (p at or below 0.01), and accuracy I-squared = 65% (p at or below 0.01). This high heterogeneity indicates that differences in study results cannot be attributed solely to sampling variability, but reflect true methodological and population differences. This is the primary reason random-effects rather than fixed-effects models were applied.
The QUADAS-2 bias assessment revealed systematic and significant concerns across all four evaluated domains. In domain 1 (patient selection), 60% of studies had well-documented acquisition protocols or used publicly available image databases. However, one study had a high risk of patient selection bias and two others had unclear risk. Critically, all studies used a retrospective case-control design in which the disease outcome was already known at the time the ML model was developed and tested. This retrospective case-control design introduces a high risk of selection bias because it does not reflect the patient spectrum a clinician would encounter in real-world practice, where the diagnosis is unknown and both PCNSL and GBM patients present mixed with other pathologies.
Index test bias (domain 2): In domain 2 (index tests), all studies had prior knowledge of the reference standard at the time they implemented their ML index test. This form of bias, sometimes called "incorporation bias," inflates apparent classifier performance because the training labels are derived from the same histopathological information that defines the disease categories. Only the single externally validated study (Kim et al., 2018) was considered to have reduced this source of bias, since its test cohort was collected independently from its training cohort. All other studies were assigned a high risk of bias in this domain. The authors explicitly note that future ML studies should use prospective designs and independent external validation to eliminate this problem.
Reference standard and flow/timing (domains 3-4): Domain 3 (reference standard) was considered adequate across studies, as histopathological confirmation of either PCNSL or GBM is a well-established and accurate gold standard. However, the authors note that detailed reporting of how biopsy specimens were obtained and processed, and the specific histological criteria used for diagnosis, was inconsistent. Domain 4 (flow and timing) was generally low-risk, as most studies clearly defined inclusion/exclusion criteria and included all eligible patients in analysis.
Radiomics Quality Score (RQS): The median RQS across all 10 studies was 16.0 out of a maximum of 36, corresponding to 44.4% of the ideal score. The range was 13 to 18 (36.1% to 50.0% of ideal). The lowest-scoring domains were high level of evidence and open science/data sharing (0% of ideal score for both), followed by biological/clinical validation and feature reproducibility. Only the externally validated study (Kim et al.) earned additional RQS points for multi-institutional validation. These RQS scores suggest that even the best-performing studies fall well below recommended best practices for radiomics research methodology, particularly regarding data sharing, prospective validation, and biological correlation.
The pooled performance metrics from this meta-analysis are encouraging, with sensitivity of 92% and specificity of 94% suggesting that ML and DL classifiers can achieve diagnostic accuracy exceeding most individual radiologists on this challenging task. The clinical implications are substantial: a reliable non-invasive classifier could allow neurosurgeons and neurooncologists to initiate appropriate chemotherapy in PCNSL patients without first performing a brain biopsy, reducing procedural morbidity, avoiding the diagnostic pitfall of steroid-induced lymphoma "melting," and potentially shortening time to treatment. For GBM patients, avoiding unnecessary empirical PCNSL-directed treatment and proceeding directly to surgical resection would also be clinically beneficial.
Comparison with prior literature: A previously published 2018 systematic review covered 8 ML-based PCNSL vs. GBM classification studies, and found similarly limited external validation, with only 1 of 8 studies externally validating their classifier. The present review, covering the 2018-2021 period, identified a total of 24 articles investigating ML and DL for this classification problem (of which 10 met strict inclusion criteria), indicating a substantial increase in research activity. The authors note a specific methodological decision that distinguishes this work: a companion paper (described as in preparation) analyzes the 12 studies comparing ML and DL against radiologist performance, while this paper focuses exclusively on the 12 studies comparing algorithms directly against histopathology as the reference standard. This separation avoids methodological conflation and allows each question to be addressed with appropriate statistical power.
Sources of heterogeneity and overfitting: The high I-squared values (65%-87%) across outcome metrics reflect genuine methodological diversity: different MRI field strengths (1.5T vs. 3T), different sequences (T1-contrast alone vs. multi-parametric), different segmentation methods (manual, semi-automatic, fully automatic), different feature extraction tools (Pyradiomics, LifeX, custom pipelines), and different classifiers (SVM, CNN, logistic regression). This heterogeneity makes meaningful pooled estimates difficult to interpret as representing any single real-world scenario. Additionally, overfitting is a recognized and pervasive problem in radiomic and deep learning studies with small training sets. Models trained on fewer than 200 cases in a single institution typically show inflated internal validation metrics that do not replicate in external cohorts.
The non-enhancing and atypical case challenge: Several studies explicitly excluded non-enhancing lesions and multiple lesions, which represent a clinically important subset of PCNSL cases, particularly in immunocompromised patients. This selective inclusion further limits the generalizability of reported metrics. Variants of PCNSL that do not show typical enhancement patterns and presentations of GBM with heterogeneous or non-enhancing components remain the most diagnostically challenging cases and are likely underrepresented in the datasets used by these studies.
Retrospective designs and case-control bias: Every one of the 10 included studies used a retrospective case-control design, meaning cases were selected because their diagnosis was already histopathologically confirmed. This creates a fundamental mismatch with clinical reality: radiologists in practice must evaluate an unselected population of patients with brain lesions where the base rate of PCNSL versus GBM versus other diagnoses is unknown. Classifiers trained on balanced or near-balanced PCNSL/GBM cohorts will have different operating characteristics in a real clinical population where GBM is far more common than PCNSL, artificially improving sensitivity at the expense of specificity in the training environment relative to practice.
Limited and potentially biased search strategy: The authors acknowledge that their search was restricted to PubMed, EMBASE, and Cochrane databases, which may have excluded relevant studies published in non-English literature, grey literature, or clinical conference proceedings. While the search covered from 1990 to December 2021, the practical body of ML and DL literature is concentrated in 2018-2021, and rapidly emerging preprints or conference papers at the time of submission may not have been captured.
Applying QUADAS-2 to ML studies: QUADAS-2 was originally designed for evaluating conventional diagnostic test accuracy studies, not ML-based classifiers. The authors note an inherent limitation in applying this tool to retrospective ML studies, particularly around the "index test" domain, where the concept of "prior knowledge of the reference standard" maps imperfectly to supervised learning, where labels are by definition known during training. Despite this limitation, the authors maintain that bias assessment remains essential even for ML studies and should become standard practice in this literature.
Absence of prospective and multicenter data: None of the 10 studies used a prospective design, and only 3 studies collected data from multiple institutions. Single-center studies using data from one scanner and one imaging protocol systematically overestimate real-world generalizability. Multi-parametric MRI protocols vary substantially between institutions, and radiomic features are known to be sensitive to acquisition parameters, scanner manufacturer, reconstruction algorithms, and slice thickness. This scan-level variability is a major barrier to deploying any of these classifiers outside their training environment without significant retraining or harmonization.
Prospective multicenter trials: The authors identify prospective multicenter trials as the most urgent need in this field. Such trials would enroll patients presenting with undiagnosed brain lesions, apply the ML or DL classifier to prospectively acquired MRI data, and then compare classifier output against subsequent histopathological confirmation as the reference standard. This design eliminates the case-control and incorporation biases that pervade current retrospective studies and would provide realistic estimates of sensitivity, specificity, and positive/negative predictive values at the actual base rate of PCNSL and GBM in the clinical population. Ideally, these trials would include diverse scanner types and field strengths to test generalizability.
Standardization of radiomics methodology: A key recommendation is the adoption of standardized pipelines for image acquisition, tumor segmentation, feature extraction, feature selection, and model validation. The Image Biomarker Standardization Initiative (IBSI) provides an established framework for radiomics feature computation that reduces inter-software variability. Professional societies including the Society of Nuclear Medicine and Molecular Imaging, the Quantitative Imaging Network, the Radiological Society of North America (RSNA), and the European Society of Radiology have all issued guidelines for radiomics research. Adherence to these guidelines, along with transparent reporting of all pipeline parameters, would substantially improve the comparability and reproducibility of future studies.
Open data and external validation as prerequisites: The 0% RQS score for open science and data sharing across all included studies highlights a major gap. Public repositories of annotated brain tumor MRI datasets would allow independent groups to train and validate classifiers on common benchmarks, accelerating progress and enabling direct comparison of competing approaches. External validation on geographically and demographically diverse cohorts, using data acquired at different institutions with different scanner hardware and protocols, should be treated as a minimum requirement before any classifier is considered for clinical deployment. Automated segmentation methods should also be validated against manual expert segmentation to establish their reliability as a component in the pipeline.
Integration with clinical workflow and non-invasive management: Beyond diagnostic accuracy, future research should evaluate whether AI-assisted MRI classification can reduce the rate of diagnostic brain biopsies in clinical practice, improve time to initiation of appropriate therapy, and ultimately improve patient outcomes. In centers where AI-supported non-invasive diagnosis becomes sufficiently reliable, it could shift clinical practice toward empirical methotrexate-based therapy trials in suspected PCNSL without prior biopsy, a strategy that has been explored in selected clinical scenarios but remains controversial. The computational infrastructure and regulatory pathways for deploying such tools in clinical settings also need to be developed in parallel with the research evidence base.