A Statistically Validated Comparison of CNN Architectures and Optimizers for Brain Tumor Classification on MRI Im-ages
International Journal of Electrical Engineering, Mathematics and Computer Science · 30.06.2026 · 3 (2) · 13-22
DOI: 10.62951/ijeemcs.v3i2.391 ↗Kurz gefasst
Fachartikel in der Zeitschrift International Journal of Electrical Engineering, Mathematics and Computer Science, online erschienen am 30.06.2026. Verfasst von I Made Dharma Yoga Pratama und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-SA 4.0).
Abstract (Originalsprache)
Brain tumors are a major global health burden, and Magnetic Resonance Imaging (MRI) is the primary modality for their detection. Manual interpretation of MRI is time-consuming and subject to inter-observer variability, motivating automated classification using Convolutional Neural Networks (CNNs). A recurring weakness in existing studies is the comparison of CNN architectures and optimizers using single-run accuracy without formal statistical testing, leaving it unclear whether reported differences are genuine or due to random training variation. This study presents a statistically validated comparison of four CNN architectures (AlexNet, VGG-16, InceptionV3, and MobileNetV2) for four-class brain tumor classification (glioma, meningioma, pituitary, and no tumor) on the Kaggle Brain Tumor MRI dataset (7,023 images). AlexNet was trained from scratch, whereas the other three used feature-extraction transfer learning with frozen ImageNet-pretrained backbones. All models were evaluated using 5-fold stratified cross-validation, and architecture differences were assessed using one-way ANOVA with Tukey HSD post-hoc testing and Cohen's d effect sizes. The best architecture was then optimized with Adam and Stochastic Gradient Descent with Momentum (SGDM), compared using a paired t-test at the 0.05 level. AlexNet achieved the highest mean test accuracy of 96.63%, significantly outperforming VGG-16 (92.60%), InceptionV3 (88.89%), and MobileNetV2 (88.25%), with large effect sizes. Adam and SGDM produced statistically equivalent accuracy (97.38% versus 96.52%). The results indicate that learning domain-specific features from scratch can surpass non-fine-tuned transfer learning, and that the choice between Adam and SGDM does not significantly affect accuracy.
Zitieren
APA
Pratama, I. M. D. Y., & Safuan, S. N. B. M. (2026). A Statistically Validated Comparison of CNN Architectures and Optimizers for Brain Tumor Classification on MRI Im-ages. International Journal of Electrical Engineering, Mathematics and Computer Science, 3(2), 13-22. https://doi.org/10.62951/ijeemcs.v3i2.391
Vancouver
Pratama IMDY, Safuan SNBM. A Statistically Validated Comparison of CNN Architectures and Optimizers for Brain Tumor Classification on MRI Im-ages. International Journal of Electrical Engineering, Mathematics and Computer Science. 2026;3(2):13-22. doi:10.62951/ijeemcs.v3i2.391
BibTeX
@article{Pratama2026,
author = {Pratama, I Made Dharma Yoga and Safuan, Syadia Nabilah Binti Mohd},
title = {A Statistically Validated Comparison of CNN Architectures and Optimizers for Brain Tumor Classification on MRI Im-ages},
journal = {International Journal of Electrical Engineering, Mathematics and Computer Science},
year = {2026},
volume = {3},
number = {2},
pages = {13--22},
doi = {10.62951/ijeemcs.v3i2.391}
}
EUVIMED