<![CDATA[Background: Deep learning studies using the widely used Brain Tumor MRI Dataset routinely report accuracies above 95%. However, the dataset combines overlapping public collections, and capsule networks are often promoted for such data without controlled comparisons. Objective: This study quantified duplicate-induced data leakage in the dataset and investigated whether capsule heads improve classification under leakage-free evaluation. Methods: Exact and near-duplicate slices were identified using file hashing and correlation analysis of head-cropped thumbnails, with the correlation threshold calibrated against label-conflict rates. Ten models using three backbones were trained under a unified protocol with five random seeds on a consumer-grade GPU. Capsule, dense, and pooling heads were compared. The models were also evaluated using the original, potentially leaky split, and their rotation robustness and inference costs were assessed. Results: Near-duplicates linked 1,707 of 5,712 training images (29.9%) to images in the test set, and data leakage inflated test accuracy by 3.22–6.33 percentage points through both memorization and the larger effective training set. With a lightweight backbone trained from scratch, the capsule head improved macro-F1 by 7.16 percentage points over the dense head while using 91.3% fewer parameters, and by 6.63 percentage points over the pooling head. These improvements were consistent across all five random seeds. However, the capsule head provided no improvement with ImageNet-pretrained backbones and did not consistently reduce the loss in accuracy under rotation. Conclusion: Capsule heads appear to offer the greatest benefit when features are learned from limited data. Leakage-free evaluation across multiple random seeds is essential for credible brain tumor MRI benchmarks. These conclusions apply to the tested public dataset and experimental protocol and do not constitute clinical validation.]]>
