Web Platform Puts Three Deep Learning Models to Work on Brain Tumor MRI Scans

Researchers have built a web-based artificial intelligence platform that lets users upload a brain MRI scan and receive an instant prediction of whether it shows a glioma, a meningioma, a pituitary tumor, or no tumor at all. The system, described in BMC Biomedical Engineering, packages three different deep learning classifiers—a custom convolutional neural network (CNN) and two pretrained architectures, VGG19 and Xception—behind a single browser interface, so that models which normally live inside research code can be compared side by side without any programming expertise.

The team, led by Jennifer Hamalaya and colleagues including medical physicists Hadi Fayad, Tarraf Torfeh and Sandy Rihana, trained and benchmarked all three models on a publicly available Kaggle dataset of 7,023 T1-weighted contrast-enhanced brain MRI images. The collection is divided into a training subset of 5,712 images and a testing subset of 1,311 images, spanning the four diagnostic classes. Each tumor class contributes more than 1,600 images and the no-tumor class contributes 2,000, a near-balance that reduces bias toward any single diagnosis.

The performance results were striking. VGG19, a 19-layer network pretrained on the ImageNet database and used here as a frozen feature extractor with a custom classification head, achieved the best test accuracy at 98 percent, with precision, recall and F1-score also at 98 percent. The custom CNN, built from scratch with five convolutional layers and a 512-neuron dense layer, reached 96.03 percent, while Xception, an architecture built on depthwise separable convolutions, achieved 96 percent. In repeated hold-out validation applied to the CNN baseline—ten independent runs with fresh random splits—the model averaged 96.5 percent accuracy with a standard deviation of 0.53 percent, indicating stable performance across randomized partitions.

Confusion matrices revealed a consistent pattern across all three models: the hardest distinction was between gliomas and meningiomas. The CNN misclassified 10 glioma samples as meningioma and 18 meningioma samples as glioma, while Xception confused 33 glioma cases with meningioma. VGG19 fared far better, with only 4 glioma-to-meningioma and 6 meningioma-to-glioma errors, and nearly perfect performance on the no-tumor class, where just one image out of 405 was misclassified. The authors attribute the glioma–meningioma overlap to the limits of working from single contrast-enhanced slices: both tumor types can appear as enhancing masses with irregular margins, and separating them reliably in the clinic typically requires additional sequences such as T2 or FLAIR that capture edema and infiltration patterns.

The experiments also produced two instructive side findings. First, preprocessing mattered in an architecture-dependent way. Without the standardized resizing and normalization pipeline, Xception’s accuracy collapsed from 96 percent to 88.59 percent, while the CNN barely moved, dropping only from 96.0 to 95.7 percent, and VGG19 slipped modestly from 98 to 97 percent. The authors link Xception’s sensitivity to its reliance on ImageNet-pretrained feature distributions. Second, data augmentation, often assumed to help, actually hurt. Under augmented training, the CNN and Xception maintained test accuracies of 91.15 and 94.28 percent respectively, but VGG19 fell to 86.80 percent. An ablation study isolating individual transformations showed that brightness adjustment and horizontal flipping were the most beneficial single augmentations, while combining all transformations simultaneously produced the worst results for every model, presumably by introducing excessive variability and a distribution shift that the frozen VGG19 backbone in particular could not adapt to.

On the engineering side, the platform follows a two-tier client–server design. A React frontend handles image upload, model selection and display of predictions with confidence scores, while a FastAPI backend performs server-side preprocessing, loads the selected model from memory, and returns the predicted class. A central model manager registers each available network, enforces its required input shape—224 by 224 pixels for the CNN and VGG19, 299 by 299 for Xception—and routes inference requests accordingly. Because models sit behind standardized endpoints, a newly trained network can be added to the registry without rewriting the application, and future modules such as segmentation or explainability tools can be slotted into the same workflow. Uploaded images are processed in memory and not permanently stored, and the interface was internally validated by the authors, who are medical physicists.

The motivation comes from a well-documented translational gap. Deep learning models for brain tumor MRI classification routinely report benchmark accuracies above 95 percent, yet clinical adoption remains limited, and only a handful of FDA-cleared AI tools target brain tumor analysis directly, among them VBrain for tumor auto-contouring and NeuroQuant Brain Tumor for metastatic and meningioma quantification. The authors also point to the subjective nature of conventional MRI interpretation: reported inter-reader agreement in brain tumor reporting sits around 80 to 82 percent, and neuroradiologists spend a mean of roughly 11 to 18 minutes interpreting each brain MRI depending on experience. A web interface that delivers automated second-opinion classifications in real time could, in principle, support radiologists at the decision-making stage rather than replace them.

The researchers are careful, however, about what these numbers do and do not establish. The dataset carries no patient demographic information, tumor grade, size or anatomical location metadata, and slice selection criteria are not described. Because no patient identifiers are included, it is impossible to confirm whether slices in the training and test sets come from distinct patients, so train–test contamination cannot be ruled out and the reported accuracies may be somewhat overestimated. The data are also single-slice, single-center and limited to a single contrast-enhanced sequence, whereas clinical diagnosis relies on multi-sequence, whole-volume MRI. Class distributions are more balanced than true clinical prevalence, and the curated Kaggle images may not capture rare tumor types, atypical presentations or acquisition artifacts. The authors explicitly state that the results should be interpreted as a demonstration of the platform and deployment framework rather than a claim of clinical performance.

Several further caveats frame the work. Repeated validation was performed only for the baseline CNN due to computational cost, leaving the transfer-learning models evaluated on a single fixed split, and no formal hyperparameter sensitivity analysis was conducted. No comparison with radiologist performance, formal usability testing or cost-effectiveness analysis was carried out, and scalability or load testing was not performed, so the current implementation is a single-user research prototype rather than a production-ready system. Explainability modules such as Grad-CAM, along with clinician-in-the-loop evaluation of the resulting saliency maps, are planned for future work, as are patient-level and multi-center external validation, multi-sequence and three-dimensional inputs, and progressive fine-tuning of the VGG19 backbone to better handle augmented data.

Even with those limits, the study offers a concrete template for moving brain tumor classifiers out of notebooks and toward clinical settings. Real deployment, the authors note, would require integration with hospital systems such as PACS and RIS, support for DICOM images and HL7 messaging, regulatory clearance such as FDA 510(k) or CE marking, and external validation with patient-level splits across multi-center datasets. The platform is positioned as a decision-support aid—with final diagnostic responsibility remaining with the clinician—for second opinions, workflow triage, and research or educational use. By exposing multiple architectures within one flexible, testable web environment, the work illustrates how AI-based neuro-oncology tools can be made more transparent and extensible, providing a stepping stone toward prospective, multi-center trials of human–AI interaction in brain tumor imaging.

Subject of Research: A web-based AI platform integrating multiple deep learning models for brain tumor MRI classification

Article Title: A web modular AI interface for brain tumor MRI classification using deep learning models

Article References: Hamalaya, J., Fayad, H., Torfeh, T., & Rihana, S. (2026). A web modular AI interface for brain tumor MRI classification using deep learning models. BMC Biomedical Engineering, 8(1), Article 19. https://doi.org/10.1186/s42490-026-00123-w

Image Credits: AI Generated

DOI: 10.1186/s42490-026-00123-w

Keywords: brain tumor, MRI, deep learning, VGG19, Xception, convolutional neural network, web platform, medical imaging, transfer learning, clinical decision support, glioma, meningioma