Evaluating transfer learning for brain tumor MRI classification: frozen VGG16 feature extraction versus end-to-end CNN training

Purpose To quantitatively compare end-to-end training of a convolutional neural network (CNN) with transfer learning using a frozen VGG16 feature extractor for multiclass brain tumor classification on contrast-enhanced magnetic resonance (MR) images. Methods A publicly available dataset of 6,056 contrast-enhanced T1-weighted brain MRI images (glioma, n = 2,004; meningioma, n = 2,004; and brain tumor, n = 2,048) was preprocessed (resized to 224 × 224, intensity-normalized, pseudo-RGB) and stratified (70/15/15 split). Five independent runs with distinct random seeds (42, 123, 2024, 7, 999) compared (1) a custom CNN trained end-to-end and (2) a frozen VGG16 with a task-specific head under identical conditions. Performance was assessed via accuracy, precision, recall, F1 score, ROC-AUC, and generalization gap (Δ_gen = training–testing accuracy) with paired t tests and Cohen’s d. Results The frozen VGG16 model achieved a mean test accuracy of 94.4% ± 0.4% (95% CI ± 0.35%), significantly outperforming the baseline CNN (58.8% ± 25.0%; paired t test p = 0.032; Cohen’s d = 1.45). VGG16 showed a near-zero generalization gap (Δ_gen = − 0.022) versus severe overfitting at baseline (Δ_gen = + 0.344). Conclusion Under strictly controlled single-dataset conditions, frozen VGG16 feature extraction establishes a statistically robust and reproducible baseline that demonstrates significant superiority over end-to-end training of a lightweight CNN (mean test accuracy 94.4% ± 0.4% vs. 58.8% ± 25.0%; paired t test p = 0.032; Cohen’s d = 1.45). External multicenter validation on diverse scanners and populations, together with quantitative interpretability analyses, remain essential prerequisites for clinical translation.