<![CDATA[Interpreting brain MRI for tumor diagnosis is time-consuming and varies between raters, and single-slice decisions can be fragile. This paper compares two classification pipelines on the Brain Tumor Classification (MRI) Dataset from Kaggle (glioma, meningioma, pituitary, no-tumor) of T1-contrast-enhanced images. The first is a per-slice CNN. The second is a per-group CRNN-GRU that reads an ordered sequence of ten slices with a bidirectional GRU and inter-slice attention. Because the dataset provides no patient identifiers, the analysis unit is a surrogate slice group, and the splits are group-disjoint by construction; all models are compared at the group-level. The per-slice CNN reached 80.0% slice-level accuracy. Aggregated to the group-level, the same CNN reached 92.68% by majority voting and 95.12% by mean-probability pooling. The CRNN-GRU reached 90.2% in a single pass, and 87.8% with test-time augmentation. No pairwise difference is statistically significant (exact McNemar, all p >= 0.25). The results do not show that sequential modelling and inter-slice attention improve on simple probability aggregation on this dataset, and the simplest configuration was the most accurate. A content-level audit found a small number of images duplicated across partitions, so all figures should be read as upper bounds. These results describe internal performance on one dataset under group-disjoint evaluation with a content-level leakage audit and do not establish clinical generalization.]]>
Copyrights © 2026
