Primary central nervous system lymphoma (PCNSL) and glioblastoma (GBM) are two of the most feared malignant brain tumors, and both carry a grim prognosis. PCNSL is a rare but aggressive non-Hodgkin lymphoma confined to the central nervous system, while GBM is the most common and lethal primary brain malignancy in adults. Although their treatments differ substantially, PCNSL being managed with high-dose methotrexate-based chemotherapy and GBM requiring surgical resection followed by concurrent chemoradiation, the two tumors can appear strikingly similar on pre-operative MRI, particularly when they involve the corpus callosum.
Why the corpus callosum is special: The corpus callosum is the large white matter bundle connecting the two cerebral hemispheres, and tumors arising here present a distinct anatomical and diagnostic challenge. Corpus callosal GBMs classically produce the "butterfly" pattern with bilateral hemispheric spread, but PCNSL can mimic this appearance. On T1-weighted gadolinium-enhanced MRI, GBM typically shows a solid homogeneous enhancement with central hypointense necrosis, while PCNSL usually demonstrates solid homogeneous enhancement without necrosis. However, pre-operative dexamethasone therapy for vasogenic edema can cause PCNSL to undergo necrosis, producing imaging features that closely resemble GBM and making visual distinction unreliable.
The biopsy risk dilemma: The definitive diagnostic approach is tissue biopsy, but corpus callosal tumor biopsy carries a reported mortality rate of 1.1% from postoperative intracranial bleeding, plus additional risks of neurological deficits. This mortality risk has motivated the search for reliable non-invasive pre-operative diagnostic tools, as correctly identifying the tumor type before surgery would allow the appropriate treatment to begin sooner and potentially spare some patients from a high-risk procedure. Prior work applying classical machine learning (ML) to this problem required manual feature extraction from images before model development, a labor-intensive semi-autonomous process that limited scalability.
The current study, conducted at Prince of Songkla University in Thailand, evaluated whether deep learning (DL) models based on convolutional neural networks (CNNs) could outperform this semi-autonomous ML paradigm by automatically learning to extract and integrate diagnostic features from T1-Gd MRI scans and reliably discriminate PCNSL from GBM in corpus callosal tumors specifically.
This is a retrospective historical cohort study enrolling patients treated between January 2010 and December 2021, a 12-year collection period designed to accumulate sufficient numbers of both tumor types. Starting from 283 patients with pathologically confirmed diagnoses of either PCNSL or GBM, the study excluded 9 individuals due to missing MRI scans or movement artifacts that compromised image quality, arriving at a final analytical cohort of 274 patients: 94 (34.3%) with PCNSL and 180 (65.7%) with GBM. All diagnoses were confirmed by a pathologist from tissue specimens, providing a solid gold standard against which DL model predictions were ultimately compared.
Corpus callosal versus non-corpus callosal cases: Within the 274-patient cohort, 46 tumors (16.8%) were located in the corpus callosum and the remaining 228 were in other brain regions. Notably, 56.5% of corpus callosal tumors were PCNSL versus only 43.5% that were GBM, a different distribution from the overall cohort where GBM predominated at 65.7%. This inversion is clinically important: PCNSL has a predilection for periventricular and corpus callosal locations, meaning this anatomical site concentrates the very diagnostic uncertainty the study aims to resolve.
Steroid use and central necrosis: Pre-operative dexamethasone was administered in 70.1% of all patients, and central necrosis was observed in 67.9% overall. Among corpus callosal tumors specifically, central necrosis was present in 43.5% of cases, with 85% of corpus callosal GBMs showing necrosis compared to only 11.5% of corpus callosal PCNSLs. This asymmetric distribution of necrosis between the two tumor types within the corpus callosum is the radiological signal that the DL model must learn to exploit, even in the presence of steroid-induced necrosis that blurs this boundary.
Image collection and pooling: Axial T1-weighted gadolinium-enhanced (T1-Gd) MRI scans were used exclusively. The total image pool comprised 1,000 MRI slices. The 228 corpus callosal MRI images were reserved as a 50/50 validation set (114 images for each tumor type in the validation dataset). The remaining 772 non-corpus callosal images were randomly split: 709 images (80%) formed the training dataset and 177 images (20%) constituted the testing dataset. This design means the DL model was trained on non-corpus callosal images and then validated on corpus callosal images it had never encountered, a rigorous out-of-distribution test of generalizability.
The DL architecture chosen for this study is a custom convolutional neural network (CNN) comprising three consecutive convolutional-pooling layers followed by a fully connected classification layer. The three convolutional-pooling layers act as hierarchical feature extractors: early layers learn low-level features like edges and contrast gradients in the MRI images, while deeper layers integrate these into higher-order representations corresponding to tumor morphology, enhancement patterns, and necrosis characteristics. The rectifier linear unit (ReLU) was used as the activation function throughout, and a batch size of 32 was employed for stochastic gradient descent. The Adam optimizer with an initial learning rate of 0.001 managed weight updates across 30 training epochs, at which point all models had reached a plateau in their training loss curves.
Model 1 (Baseline): The first model was trained without any regularization or data augmentation, serving as an unmodified baseline to establish the CNN's intrinsic classification ability on this dataset. This model overfits in a detectable way: the testing loss rose away from the training loss as epochs progressed, indicating the model was memorizing the training data rather than generalizing.
Model 2 (L2 Regularization): Ridge regression, also known as L2 regularization, was applied in the second model. L2 regularization adds a penalty proportional to the squared magnitude of the model weights to the loss function, discouraging the network from assigning excessively large weights to any individual feature and thereby reducing overfitting. Model 2 achieved testing accuracy of 0.72, compared to 0.69 for the baseline.
Model 3 (Dropout Regularization): The third model used dropout regularization with a rate of 0.25, meaning 25% of neurons were randomly deactivated during each training pass. Dropout forces the network to learn redundant representations and prevents co-adaptation of neurons, a different mechanism from L2 that also reduces overfitting. Model 3 reached testing accuracy of 0.73. A comparison of Models 2 and 3 showed that Model 3's testing loss was higher than Model 2's as training proceeded, suggesting L2 provided slightly better stability.
Model 4 (Data Augmentation): The fourth model applied data augmentation using random vertical flips, horizontal flips, rotations, and zoom operations to artificially expand the training dataset. Despite the theoretical benefit of reducing variance through increased training diversity, Model 4 actually performed worst, with testing accuracy of 0.61 and the lowest validation accuracy of 0.68 among the four models. All models were developed and validated using Python with Keras version 2.4.0.
The diagnostic performance of all four models was formally evaluated on the unseen corpus callosal validation dataset of 114 MRI images, providing out-of-distribution test results that reflect how the models would perform on the clinically challenging cases. Multiple performance metrics were computed: sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), overall accuracy, F1 score, and AUC. PCNSL was treated as the positive class (class 1) and GBM as the negative class (class 0), so sensitivity measures the model's ability to correctly identify PCNSL and specificity measures its ability to correctly identify GBM.
Model 1 (Baseline) results: On the validation corpus callosal images, the baseline model achieved sensitivity of 0.59 (95% CI: 0.46-0.72), specificity of 0.94 (95% CI: 0.88-1.00), PPV of 0.91 (95% CI: 0.83-1.00), NPV of 0.70 (95% CI: 0.59-0.80), accuracy of 0.77 (95% CI: 0.69-0.84), F1 score of 0.72 (95% CI: 0.59-0.85), and AUC of 0.77. The high specificity with moderate sensitivity indicates the model was conservative in calling PCNSL, more often misclassifying PCNSL as GBM than the reverse.
Model 2 (L2) and Model 3 (Dropout) results: Both regularized models improved substantially over the baseline. Model 2 achieved sensitivity of 0.82 (95% CI: 0.72-0.92), specificity of 0.84 (95% CI: 0.74-0.93), PPV of 0.83 (95% CI: 0.74-0.93), NPV of 0.82 (95% CI: 0.73-0.92), accuracy of 0.83 (95% CI: 0.76-0.90), F1 of 0.83 (95% CI: 0.73-0.93), and AUC of 0.83. Model 3 reached sensitivity 0.75, specificity 0.92, PPV 0.91, NPV 0.79, accuracy 0.84, F1 0.82, and AUC of 0.84. Model 3 had the highest AUC at 0.84, though both Models 2 and 3 represent high discrimination compared to the acceptable level of Model 1.
Model 4 (Data Augmentation) results: Despite the appeal of augmentation for small datasets, Model 4 performed worst: sensitivity 0.75, specificity 0.61, PPV 0.66, NPV 0.71, accuracy 0.68, F1 0.70, AUC 0.57. An AUC of 0.57 is barely above the no-discrimination threshold of 0.50, indicating that data augmentation with random flips, rotations, and zoom actually degraded model quality for this specific MRI classification task. One plausible explanation is that MRI image orientation carries diagnostic meaning that flipping and rotation destroys, making the augmented training data misleading rather than helpful.
Following model validation, the best-performing model was deployed as a clinical web application to facilitate real-world use. Model 2 (L2 regularization), rather than Model 3 (dropout), was selected for deployment despite Model 3 having a marginally higher AUC of 0.84 versus 0.83. The rationale was that Model 2 demonstrated a lower testing loss relative to Model 3 as training epochs progressed, suggesting more stable convergence and slightly better calibration, which is an important property for a tool that clinicians will rely on for individual patient predictions.
Technical implementation: The web application was built using Streamlit version 0.74.1, an open-source Python framework designed specifically for rapid deployment of data science tools as interactive web interfaces. Streamlit requires no web development expertise beyond Python, making it an accessible deployment pathway for research groups without dedicated software engineering resources. The application accepts new T1-Gd MRI images uploaded via computer or mobile phone, processes them through the trained CNN, and outputs a predicted diagnosis (PCNSL or GBM) with associated probability.
Access and workflow: The tool was designed for ease of clinical use, accessible by scanning a QR code or entering a URL directly into any browser. This accessibility is important for the target clinical context of this tool, which is rural or resource-limited neurosurgical settings where specialist neuroradiologists may not be available to interpret ambiguous MRI findings in real time. The authors envision this tool as a screening aid that flags cases needing expedited specialist review or guides empirical management when immediate biopsy carries unacceptable surgical risk.
Telemedicine integration potential: The study explicitly positions this web application within a telemedicine and teleconsultation framework, citing prior work by Wong and Tsai on DL-based telehealth for glaucoma screening. The logic is that a lightweight, mobile-accessible diagnostic classifier could extend specialist-level imaging interpretation to hospitals in regions with limited neuroradiology infrastructure, reducing diagnostic delays and the risk of treating PCNSL patients with GBM protocols or vice versa, a clinically consequential error given the fundamentally different first-line therapies for the two diseases.
Several prior studies provide important context for interpreting the AUC values of 0.83-0.84 achieved here. McAvoy et al. applied a CNN to classify PCNSL and GBM from MRI and reported AUC values of 0.94-0.95, considerably higher than the present study's best model. Zhang et al. extended DL to a three-class problem differentiating PCNSL, GBM, and tumefactive demyelinating lesions, achieving diagnostic performance between 0.95 and 1.00. Tariciotti et al. developed a DL model for preoperative differentiation of GBM, brain metastasis, and PCNSL, and Xia et al. trained a multi-parametric MRI-based CNN for PCNSL-GBM differentiation. These high-performing studies generally used multiple MRI sequences (not just T1-Gd) and did not restrict their analysis to the notoriously ambiguous corpus callosal location.
Traditional ML comparison: Classical supervised ML approaches using support vector machines (SVM) and multilayer perceptrons for PCNSL-GBM classification achieved a maximum accuracy of 69.2%. A systematic review and meta-analysis by Nguyen et al. of ML studies for this classification task found reported AUC values ranging from 0.878 to 0.979, with SVM being the most frequently applied algorithm. The present study's best DL AUC of 0.84 falls below this ML meta-analysis range, but it is important to note that those ML studies required expert radiomics feature extraction as a preprocessing step, while the DL approach here automates feature extraction entirely from raw image data.
Apparent diffusion coefficient (ADC) mapping: Horger et al. evaluated ADC maps, derived from diffusion-weighted MRI, for distinguishing primary CNS lymphoma from astrocytic tumors infiltrating the corpus callosum, achieving sensitivity of 73-84% and specificity of 89-100%. ADC captures tissue microstructural properties based on water diffusion restriction, and PCNSL characteristically shows marked diffusion restriction due to its high cellularity, in contrast to GBM where necrosis produces heterogeneous diffusion patterns. The present study did not incorporate diffusion-weighted sequences, which may partly explain the performance gap relative to those methods.
In aggregate, the present study sits in a lower-performance range than many prior PCNSL-GBM DL studies, but occupies a distinct niche: it specifically addresses corpus callosal tumors, uses only a single MRI sequence (T1-Gd), and trains on a dataset derived from a single Southeast Asian tertiary center. These constraints make direct comparison difficult but also highlight the gap between what DL can achieve with rich multi-sequence imaging and what is achievable in the more constrained clinical scenarios the authors are targeting.
Single MRI sequence: This study used only axial T1-weighted gadolinium-enhanced images for DL model development. Modern brain tumor MRI protocols routinely include additional sequences such as T2-weighted FLAIR, diffusion-weighted imaging (DWI) with ADC maps, dynamic susceptibility contrast perfusion, and MR spectroscopy. Each additional sequence provides orthogonal biological information: DWI captures cellularity differences, perfusion imaging assesses tumor vascularity, and spectroscopy measures metabolite profiles. The authors explicitly acknowledge that incorporating multiple MRI sequences would likely enhance diagnostic performance in future iterations, a conclusion supported by the substantially higher AUCs reported in multi-parametric studies.
Limited training data: The total dataset of 1,000 MRI images, with only 114 unseen corpus callosal images used for validation, is small by DL standards. The authors cite the widely referenced heuristic that robust image classification with DL typically requires 1,000 images per class, meaning the current dataset falls substantially short for training on PCNSL alone (94 patients total in the cohort). Small training sets increase the variance of model estimates and raise the risk that reported AUC values may not be reproducible at other institutions or on different patient populations. The class imbalance (94 PCNSL vs. 180 GBM) further complicates model training, as standard cross-entropy loss may bias the model toward predicting the majority class.
Single-center retrospective design: All imaging data originated from a single academic medical center in southern Thailand. Scanner hardware, imaging protocols, gadolinium dosing, and patient population characteristics at Prince of Songkla University may differ from those at other institutions. Retrospective designs are also subject to selection bias, as only patients who eventually received tissue diagnosis are included, excluding cases where clinical or imaging findings were sufficiently characteristic to guide treatment without biopsy. External validation at independent centers is absent from this study.
Data augmentation failure: The unexpected performance drop with data augmentation (AUC falling to 0.57) warrants cautious interpretation. Random flips and rotations may corrupt the spatial orientation cues that carry diagnostic meaning in brain MRI, where left-right and anterior-posterior asymmetries are clinically meaningful. More domain-appropriate augmentation strategies, such as MRI-specific noise injection, contrast variation, or realistic deformation fields, might avoid this pitfall in future work.
Multi-parametric MRI integration: The most straightforward performance improvement would come from incorporating additional MRI sequences alongside T1-Gd. DWI and ADC maps are particularly promising given that PCNSL reliably shows greater diffusion restriction than GBM due to its high nuclear-to-cytoplasm ratio and dense cellularity. Multi-parametric CNN architectures can accept multiple image channels simultaneously, analogous to how standard CNNs process RGB color channels, and several published studies have demonstrated AUC improvements of 0.05-0.15 when transitioning from single-sequence to multi-parametric DL models for brain tumor classification.
Multi-center data pooling: Addressing the training data shortage requires collaboration across multiple institutions. Multi-center studies would provide more images of corpus callosal tumors specifically, reducing the domain mismatch between the training set (non-corpus callosal images) and the validation set (corpus callosal images) that the authors had to accept due to sample size constraints. Federated learning, which trains models across institutions without sharing raw patient data, offers a privacy-preserving pathway to achieve this scale for rare intracranial tumors like corpus callosal PCNSL and GBM.
Transfer learning and pre-trained architectures: The current study used a relatively shallow custom CNN trained from scratch. Applying transfer learning from architectures pre-trained on large medical imaging datasets, such as ResNet50, EfficientNet, or vision transformers adapted for radiological images, could substantially improve performance, particularly when the labeled dataset is limited. Pre-trained models have already learned generalizable low-level image features and require fine-tuning only on the task-specific classification head, reducing the amount of labeled data needed to achieve competitive performance.
High-sensitivity screening tool integration with telemedicine: The authors envision this approach fitting into a triage workflow where the DL classifier, running as a mobile-accessible web application, flags ambiguous corpus callosal cases for teleconsultation with neuroradiology or neuropathology specialists at central academic centers. For rural hospitals where MRI is available but specialist review is delayed by days, a high-sensitivity DL classifier that rarely misses PCNSL (even at the cost of some false positives) could accelerate time to appropriate treatment. High-sensitivity tools are particularly valuable in this context because the consequence of initiating GBM-directed steroids and radiation in an undiagnosed PCNSL patient includes masking tumor on subsequent MRI and delaying potentially curative chemotherapy.