Ensemble AI Reads Brain MRI With Adaptive Attention and Statistical Rigor

Researchers have built a deep learning model that classifies brain tumors in MRI scans across four categories—glioma, meningioma, pituitary tumor, and no tumor—while pairing its accuracy with formal statistical testing and visual explanations of its decisions. The system, called CAE-BrainNet, is described in a study published in the journal Brain Informatics by a team led by Mehedi Hasan Meraj and colleagues, including Kamanashis Biswas.

The model is an ensemble: it combines three convolutional neural networks chosen for their architectural differences. EfficientNetV2-M uses compound scaling with inverted residual blocks, DenseNet201 relies on dense connectivity to reuse features across layers, and ConvNeXt-Base modernizes the ResNet design with depthwise convolutions. The authors argue that these three bring complementary inductive biases, meaning each network is naturally suited to different morphological patterns in tumor tissue. In their evaluation, transformer-based alternatives such as ViT-B/16, Swin Transformer V2, and MaxViT-Tiny performed worse on the benchmark and were excluded from the ensemble.

The distinctive component is a class-adaptive attention mechanism. Instead of averaging the three networks’ outputs with fixed weights, the model learns a four-by-three weight matrix in which each entry governs how much a given backbone contributes to a given tumor class. A trainable temperature scalar controls how sharply the attention concentrates: small values push the model to trust one network strongly, while large values spread the contribution more evenly. A preliminary probability estimate, formed by averaging the three models’ predictions, feeds this attention module, and the resulting weighted features pass through a fusion network that produces the final four-class output.

Setting the weights was itself a two-stage process. A coarse-to-fine grid search over a constrained weight simplex provided a robust initialization, avoiding the sparse, zero-weight solutions that purely discrete searches tend to produce. Gradient-based refinement then softened those weights during ensemble fine-tuning with five-fold stratified cross-validation. The authors report that an ablation study supports this design: removing the class-adaptation lowered accuracy to 98.78 percent, and fixing the temperature to one lowered it to 99.16 percent, compared with 99.39 percent for the full model.

The training data combined three public MRI repositories into a set of 7,023 T1-weighted contrast-enhanced images: 1,621 glioma, 1,645 meningioma, 1,757 pituitary tumor, and 2,000 no-tumor scans. The sources were the Figshare dataset compiled by Cheng and colleagues, the SARTAJ collection—whose glioma images the authors replaced with verified Figshare samples after preliminary checks revealed systematic mislabeling—and the Br35H dataset for healthy scans. The data were split 81:19 into training and test sets, with the 1,311-image test set held out until final evaluation, and a fixed random seed used throughout.

Preprocessing proceeded in four stages: automated extraction of brain tissue using Otsu thresholding and morphological operations, contrast enhancement with CLAHE, resizing to 224 by 224 pixels with architecture-specific normalization, and augmentation applied only to training images. On the held-out test set, CAE-BrainNet reached 99.39 percent accuracy, correctly classifying 1,303 of 1,311 scans, with eight errors distributed across glioma, meningioma, and pituitary cases. The model achieved 100 percent recall on the no-tumor class and 99.89 percent specificity in identifying healthy tissue, a metric the authors highlight because false positives in this category can trigger unnecessary follow-up.

A key feature of the study is its statistical validation, which the authors say is rare in this literature. McNemar’s test, applied to paired predictions, showed significant gains over the base models—for example, against EfficientNetV2-M the comparison yielded a chi-square statistic of 22.13 with p below 0.0001 and a Cohen’s g effect size of 0.39. Cochran’s Q test across models rejected the hypothesis of equal performance with Q equal to 30.11 on three degrees of freedom and p below 0.0001. Wilson score intervals quantified uncertainty, and the authors note one nuance: the comparison with ViT-B/16 was not statistically significant in aggregate, though the ensemble corrected 19 of that model’s errors while being corrected on 11.

To test generalization beyond the training distribution, the models were evaluated zero-shot—without any retraining—on the independent BRISC2025 dataset, a collection of radiologist-annotated scans aggregated from multiple sources. CAE-BrainNet scored 98.10 percent there, a drop of only 1.29 percentage points from its benchmark result, and outperformed all CNN baselines and MaxViT-Tiny, with McNemar tests and effect sizes supporting the differences. The authors interpret this as evidence that the attention mechanism learned tumor morphologies that transfer across scanner and institutional variation, though they present it as a benchmark result rather than a clinical trial finding.

Explainability received quantitative treatment as well. Grad-CAM heatmaps were computed for each backbone and for the ensemble as a whole, and the spatial overlap between models’ attention was measured with intersection-over-union scores. The mean IoU was modest at 0.16, which the authors frame as a signature of complementary rather than redundant attention: networks focus on different regions, such as tumor cores versus surrounding tissue. They also introduce AWLCF, an agreement-weighted variant of LIME that filters explanations by cross-model consistency, producing sparser, anatomically localized saliency maps. The full inference pipeline, including Grad-CAM generation, took 16.58 seconds per image, and the model is deployed as an open web application on Hugging Face for independent testing.

Subject of Research: A deep learning ensemble model for explainable, statistically validated classification of brain tumors from MRI scans

Article Title: CAE-BrainNet: a statistically validated class-adaptive attention ensemble model for explainable brain tumor classification from MRI

Article References: Meraj, M. H., Tajbid, M., Mojumdar, M. U., Chakraborty, N. R., Chowdhury, M. J. M., & Biswas, K. (2026). CAE-BrainNet: a statistically validated class-adaptive attention ensemble model for explainable brain tumor classification from MRI. Brain Informatics, 13(1), Article 40. https://doi.org/10.1186/s40708-026-00325-x

Image Credits: AI Generated

DOI: 10.1186/s40708-026-00325-x

Keywords: brain tumor classification, MRI, deep learning, ensemble model, attention mechanism, Grad-CAM, explainable AI, statistical validation, McNemar test, Cochran's Q test, zero-shot transfer, convolutional neural networks