Addressing challenges in accurate brain tumor classification in MRI: a transfer learning approach with EfficientNetB3 and comprehensive model evaluation

Brain tumors pose significant health risks because of uncontrolled and abnormal cell growth, potentially leading to the devastating of certain organs and even death, mostly in adult populations. So, early and precise brain tumor classification (BTC) is essential, as a timely diagnosis directly impacts the efficacy of treatment and the survival rates of patients. Magnetic Resonance Imaging (MRI) is the preferred diagnostic tool for BTC due to its superior image quality, yet manual classification methods have the disadvantage of being time-consuming and labour-intensive. The maturity of deep learning (DL) in modern times has brought remarkable progress in automating complex processes across multiple domains, including transfer learning (TL) in medical imaging. In our study, we propose an approach to partially automate the BTC task through the application of transfer learning on MRI data. After applying image pre-processing techniques like cropping, enhancement and Augmentation, we tested six architectures - EfficientNetB6, EfficientNetB3, VGG19, NASNetLarge, DenseNet169 and DenseNet201. Our experiments on a comprehensive publicly available kaggle dataset show that the EfficientNetB3 model achieved the best performance, with a training accuracy of 100% and test accuracy of 99.62%, outperforming all other architectures. EfficientNetB6 demonstrated a similar accuracy of 99.62% , while DenseNet201 and NASNetLarge performed well, with accuracies of 99.24% and 99.08% , respectively. VGG19, although a popular baseline model, achieved a lower accuracy of 94.96%. These findings suggest that EfficientNetB3, alongside other advanced architectures, offers a robust and accurate solution for brain tumor classification in MRI, paving the way for more reliable and efficient diagnostic tools in clinical settings.