The problem the team set out to address is deceptively simple but has far-reaching consequences. Brain tumors remain a major global health challenge, and magnetic resonance imaging is the most commonly used and reliable modality for early detection. Over the past decade, deep learning has transformed medical image analysis, and datasets hosted on platforms such as Kaggle have become de facto standards for training and comparing classification models. Yet when the same scan, or near-identical crops and resampled versions of the same scan, appear in both the training and test sets, a model can memorize specific images rather than learn generalizable pathological features. The result is a test score that looks spectacular but collapses when the model confronts genuinely new patients. Earlier work in the field, including calls for transparency published in Nature Medicine and analyses of leakage in machine learning-based science, has warned that this phenomenon distorts the literature, but few studies have quantified its impact on brain tumor MRI benchmarks or provided a practical remedy.

The Northeastern team, led by Dong Lu, Yu Zhang, and Divya Chaudhary, began by systematically analyzing a widely adopted public dataset and developing an automated data cleaning pipeline capable of identifying and removing duplicate scans. Rather than relying on visual inspection, the pipeline detects images that are duplicates or near-duplicates, which is essential because exact file matching would miss rescaled, re-cropped, or lightly compressed copies of the same underlying scan. Applying this pipeline to the popular dataset produced a leakage-free benchmark containing 3,522 unique MRI scans, covering the standard categories used in brain tumor classification, including glioma, meningioma, pituitary tumors, and healthy tissue. All comparative experiments in the new study were run exclusively on this cleaned dataset, so every model was evaluated on images it had genuinely never seen during training.

With a trustworthy benchmark in place, the researchers turned to the model architecture itself. Their starting point was the Swin Transformer, a hierarchical vision transformer that has become one of the most influential architectures in computer vision since its introduction at the IEEE/CVF International Conference on Computer Vision in 2021. Unlike the original Vision Transformer, which treats an image as a flat sequence of 16-by-16 pixel patches and applies global self-attention at every layer, the Swin Transformer organizes computation into local windows and shifts those windows between successive layers, allowing information to flow across the image while keeping computational cost linear in the number of pixels. This hierarchical, windowed design produces multi-scale feature maps much like a convolutional neural network, which is one reason it has proven popular for medical imaging tasks where lesions appear at very different scales.

The new contribution, MTA-Swin, re-designs the attention computation within Swin blocks by incorporating what the authors call Multi-Token Attention, an idea drawn from recent research on attention mechanisms in large language models. In a standard transformer, attention logits are computed between individual query and key vectors, meaning each token attends to other tokens one at a time. Multi-Token Attention instead refines the attention logits so that groups of nearby tokens can jointly influence the attention distribution, enriching the local context available to each attention head. In effect, the model can recognize that a cluster of tokens representing a tumor boundary carries more meaning together than any single token does in isolation. The Northeastern implementation goes further by enabling explicit cross-head information exchange in deeper layers of the network, allowing different attention heads, which typically operate independently and may capture complementary features, to share what they have learned. Crucially, all of this is achieved while preserving the Swin Transformer’s hierarchical stages and windowing scheme, so the architecture retains its computational efficiency and its ability to build multi-scale representations.

Training followed a two-stage transfer learning strategy. MTA-Swin was first pre-trained on ImageNet-1K, the large-scale natural image dataset that has long served as the starting point for visual representation learning, and was then fine-tuned on the leakage-free brain tumor dataset. Transfer learning of this kind is standard practice in medical imaging, where labeled clinical data is scarce, but the authors note that the choice of benchmark makes all the difference in how much of the reported performance is real. The team also examined an information bottleneck inherent in the patch tokenization process: each 4-by-4 image patch is flattened and linearly projected into a single embedding vector, a process that introduces spatial relationship loss, compressing corner features, edge patterns, and local geometric structures within the patch, and feature mixing, in which a patch containing multiple distinct textures or structures, such as a tumor boundary region spanning pathological and normal tissue, is blended into a single averaged representation that can obscure critical pathological details. These observations motivated attention designs that recover some of the lost local structure.

The headline result is striking. Across three random seeds, MTA-Swin reached an overall accuracy of 98.57 percent on the leakage-free benchmark, surpassing thirteen representative baseline models that include both convolutional neural networks and other transformer variants. Because the evaluation was performed on cleaned data, this figure represents a genuinely honest measurement of generalization, and the fact that it remains so high suggests that a well-designed architecture can indeed learn clinically meaningful features from MRI once the distorting effect of duplicates is removed. The authors went beyond a single train-test split, however. They conducted additional stratified cross-validation to confirm that the result was not an artifact of a favorable data partition, and the model’s performance held up across folds.

Interpretability received equally careful treatment. Using Grad-CAM, a technique that produces heatmap visualizations of which image regions most influenced a network’s prediction, the researchers analyzed where the model focused when classifying each tumor type. The analyses supported the robustness and interpretability of the proposed model, showing that attention concentrated on tumor-relevant regions rather than on background anatomy, scanner artifacts, or other spurious cues. This matters for clinical adoption, because a computer-aided diagnostic tool that cannot explain its decisions is unlikely to earn the trust of radiologists, and models that latch onto irrelevant features are a well-documented failure mode in medical imaging AI. Grad-CAM evidence that the network looks at the lesion itself is a meaningful step toward a practical diagnostic support system.

The broader implications of the study extend well beyond a single architecture. The authors’ automated cleaning pipeline is, in many respects, as significant as the model itself. Widely used public brain tumor MRI collections have been the foundation for hundreds of published papers, and if duplicate-induced leakage is as prevalent as this analysis suggests, a substantial fraction of reported results in the literature may need to be reinterpreted. The researchers point out that similar concerns apply across medical imaging AI, where benchmark contamination has been described as an urgent problem requiring transparency from dataset creators and model developers alike. By releasing their cleaned dataset publicly on Kaggle, the team has given the community a common, leakage-free yardstick against which future models can be measured on equal terms.

For clinicians and researchers, the study offers two practical takeaways. First, any deep learning result reported on an uncleaned public MRI dataset should be treated with caution, and groups can adopt automated duplicate-detection pipelines as a standard pre-processing step before benchmarking. Second, architectures that enrich local context within attention computations and allow information exchange between attention heads appear to offer real gains for medical image classification, suggesting a productive direction for future transformer designs in radiology. The work, published as Volume 50, article 100 of the Journal of Medical Systems, was received in October 2025, accepted in May 2026, and published on 22 June 2026, with no specific funding declared and no competing interests reported by the authors. As artificial intelligence moves closer to the radiology reading room, the Northeastern team’s message is clear: before asking how accurate a model is, first make sure the test itself is clean.