Cross-Domain Transfer Learning for Brain Tumor Classification Under Limited MRI Data Regimes: A Fraction-Resolved Benchmark with Explicit Statistical and Methodological Caveats

<title>Abstract</title> <p>Brain tumors cause an estimated 251,000 deaths worldwide each year \cite{sung2021}, yet the labeled MRI archives that deep learning classifiers typically require remain out of reach for many hospitals. We compared ImageNet-pretrained against randomly initialized EfficientNet-B0 for four-class brain tumor MRI classification across six training-data fractions (5--100\% of a 5,600-image, class-balanced pool from the Nickparvar Kaggle collection), on a fixed 1,600-image test set with three seeds per condition. Transfer learning produced a statistically confirmed improvement only under the most extreme data scarcity we tested (5\%, 238 images: +11.4 points, 95\% CI [11.0, 11.8]~pp, $p &lt; .001$); at every larger fraction the fine-tuned model scored higher on average, but the difference was not statistically distinguishable from chance under our three-seed design. We also report two minor descriptive quantities, a Transfer Efficiency Index and a Minimum Viable Data Threshold, but are explicit that neither has a theoretical derivation or independent validation. Per-class and Grad-CAM++ analyses identify glioma as the most persistently misclassified class, with attention drifting toward ''no tumor'' under uncertainty, favoring a screen-then-refer deployment over an autonomous one. Because the dataset lacks confirmed patient-level identifiers, we cannot rule out same-patient slices in both partitions, the study's most consequential limitation. We close with a concrete roadmap: more seeds, additional backbones, patient-disjoint validation, and a learning-curve alternative to TEI, all needed before any number here should inform a real deployment decision.</p>