An enhanced convolutional neural network model for brain tumor classification using transfer learning and data augmentation

Abstract

Brain tumors constitute a significant global health challenge, and accurate, timely diagnosis is critical for effective treatment planning. Manual interpretation of magnetic resonance imaging (MRI) remains time-consuming, error-prone and subject to inter-observer variability. The objective of this study was to develop and evaluate a fine-tuned VGG16 convolutional neural network (CNN), combined with a dual online and offline data augmentation pipeline, for the three-class classification of brain tumors from T1 contrast-enhanced MRI under resource-constrained computing conditions. The study used 3,064 T1 contrast-enhanced MRI slices spanning three tumor classes, namely glioma (1,426 slices), meningioma (708 slices) and pituitary tumor (930 slices), drawn from the publicly available Masoud Nickparvar Brain Tumor MRI Dataset on Kaggle. Images were normalised using min-max scaling, denoised using a Gaussian filter and resized to 224 x 224 pixels. The cohort was partitioned by stratified sampling into 2,145 training (70%), 273 validation (9%) and 646 test (21%) images. Six architectures were benchmarked, comprising a custom CNN, VGG16, VGG19, Xception, ResNet50 and DenseNet121. The best-performing backbone was then enhanced through layer freezing and fine-tuning, and finally through augmentation using random horizontal flipping applied online and histogram equalisation applied offline. The novelty of the study lies not in a new network topology but in the empirical demonstration that a selectively fine-tuned VGG16 backbone, coupled with a reproducible dual augmentation protocol, attains competitive diagnostic performance on central processing unit hardware alone. On the held-out test set, the fine-tuned VGG16 with augmentation achieved an accuracy of 98.45%, a macro-averaged sensitivity of 98.26%, a macro-averaged specificity of 99.19% and a macro-averaged precision of 98.37%. Class-wise areas under the receiver operating characteristic curve were 0.998, 0.996 and 0.999 for glioma, meningioma and pituitary tumor respectively. The custom CNN attained 95.40% accuracy and DenseNet121 attained 93.18%. These findings indicate that transfer learning based CNN models with strategic augmentation can reach diagnostically useful performance without specialised graphics hardware, offering a practical decision-support aid for radiological workflows in low-resource settings. External validation on independent clinical cohorts is required before deployment.