A Hybrid EfficientNetV2B3–ResNetV2 Deep Transfer Learning Framework for Multiclass Brain Tumor Classification in MRI

Abstract

Brain tumors are among the most life-threatening neurological disorders, and patient outcomes depend heavily on early, accurate diagnosis. Manual interpretation of magnetic resonance imaging (MRI) is time-consuming and subject to inter-observer variability, motivating automated deep-learning-based diagnostic support. This paper presents two supervised deep transfer learning pipelines for classifying brain MR images into four categories: glioma, meningioma, pituitary tumor, and no tumor. The first pipeline fine-tunes a pretrained DenseNet201 network with an attention-weighted feature map and a custom classification head. The principal contribution is a hybrid architecture that fuses deep features extracted in parallel by EfficientNetV2B3 and ResNetV2, concatenates their pooled representations, and passes them through sixteen custom layers before final SoftMax classification. Both models were trained on a merged public dataset of 7,023 T1-weighted MR images (Br35H, SARTAJ and Figshare) after grayscale conversion, denoising, contrast-limited adaptive histogram equalization (CLAHE), normalization, and geometric data augmentation. The fine-tuned DenseNet201 model attained 98.09% test accuracy, while the proposed hybrid EfficientNetV2B3–ResNetV2 model achieved a superior 99% test accuracy with a macro-averaged F1-score of 0.99, outperforming the single-backbone baseline as well as several previously reported approaches evaluated on comparable datasets. The results indicate that fusing complementary residual and compound-scaled feature representations improves multiclass brain tumor discrimination, particularly for the visually overlapping glioma and meningioma classes.