Brain Tumor Mri Eval

---
name: brain-tumor-mri-eval
description: Evaluates a model's ability to classify brain MRI scans into four pathological categories (Glioma, Meningioma, Pituitary Tumor, or None) using a hybrid CNN-ViT architecture with adaptive attention gating. Use when the user wants to benchmark on Brain Tumor MRI Dataset, or asks about evaluating this task. Reports accuracy.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2604.23137
  bibtex_key: hasnain2026cnnti
  confidence: medium
---

# brain-tumor-mri-eval

> CNN-ViT Fusion with Adaptive Attention Gate for Brain Tumor MRI Classification: A Hybrid Deep Learning Model — Hasnain et al. (2026) (arXiv:2604.23137, 2026)

## What this evaluates

Evaluates a model's ability to classify brain MRI scans into four pathological categories (Glioma, Meningioma, Pituitary Tumor, or None) using a hybrid CNN-ViT architecture with adaptive attention gating.

## Datasets

- **Brain Tumor MRI Dataset** — total 7023; splits: train (6723), test (1311)

## Metrics

- `accuracy` **(primary)** — range: [0, 1]
  - Standard classification accuracy: the proportion of correctly predicted class labels out of the total number of test instances.

## Input / output format

**Input**: Single-channel MRI images resized to 128x128 pixels, with pixel values normalized to [0, 1].

**Output**: Predicted class label from four categories: Glioma, Meningioma, None of the tumors, or Pituitary Tumor.

## Scoring recipe

```python
def compute_accuracy(predictions, gold_labels):
    correct = sum(1 for pred, gold in zip(predictions, gold_labels) if pred == gold)
    return correct / len(gold_labels)
```

## Common pitfalls

- The dataset uses a fixed train/test split provided by the Kaggle source; no cross-validation or additional random splitting was applied.
- Images are downsampled to 128x128, which may lose fine-grained pathological details compared to original high-resolution scans.
- Online augmentation is limited to mild transformations (±5% rotation/zoom, horizontal flip) to preserve anatomical plausibility, which may restrict generalization to heavily distorted scans.

## Evidence (verbatim from paper)

> The dataset was split as provided by the training/testing split; no further cross-validation-based splitting was applied to the dataset... Each image was downsampled to 128x128 pixels to have uniform input sizes and make the computationally manageable. Possible preprocessing of data involved normalizing the values in the pixel to [0, 1] by dividing by 255. Experimental results on the Kaggle Brain Tumor MRI Dataset show 97.60% test accuracy, outperforming CNN-only, ViT-only, and existing fusion methods...

## Citation

```bibtex
@misc{hasnain2026cnnti,
  title={CNN-ViT Fusion with Adaptive Attention Gate for Brain Tumor MRI Classification: A Hybrid Deep Learning Model},
  author={Hasnain et al. (2026)},
  year={2026},
  note={arXiv:2604.23137}
}
```

- arXiv: 2604.23137