Abstract
Near-ceiling accuracies are frequently reported on public brain tumor magnetic resonance imaging (MRI) classification benchmarks, yet the integrity of the underlying data partitions remains insufficiently examined. We analyze a widely used four-class dataset whose release documentation states that duplicate images and train/test overlap were removed. Although exact hashing finds no byte-identical files shared across the official partitions, a correlation-based analysis accounting for rotations and reflections identifies a near-duplicate training counterpart for 1,100 of the 1,600 test images (68.8%). On the official partition, a ResNet18 achieves an accuracy of 0.9973 on matched test images, compared with 0.8540 on unmatched images, a difference of 14.3 percentage points. We therefore construct a cluster-disjoint partition by assigning all images within each detected near-duplicate cluster exclusively to one subset. On the resulting test set, DINOv2 with low-rank adaptation (LoRA) achieves the highest observed accuracy of 0.9748 and a macro-F1 score of 0.9743 while updating only 1.02% of its parameters. Its performance is comparable to the 0.9721 accuracy of fully fine-tuned EfficientNet-B0, which requires 18 times more trainable parameters, and exceeds that of a frozen DINOv2 linear probe by 6.22 percentage points. These findings demonstrate the importance of near-duplicate-aware partitioning for credible benchmark evaluation and establish LoRA-adapted DINOv2 as an accurate, parameter-efficient baseline for brain tumor MRI classification.