Unlocking Magnetic Resonance Images Potential: Innovative Brain Tumor Classification via Transfer Learning and Convolutional Neural Networks

Background and Objectives: Recent advancements in artificial intelligence (AI), particularly through convolutional neural networks (CNNs), have led to significant progress in tumor detection and classification systems. This study aims to train and evaluate the performance of four distinct CNN models for the 4-way classification of glioma, meningioma, pituitary tumors, and non-tumor magnetic resonance images (MRI). Methods Data augmentation techniques were applied to MRI slices from the Figshare, Br35H, and Brain Tumor Classification datasets to enhance the generalizability of the models and mitigate overfitting, effectively increasing the sample size. Four pre-trained CNN models including VGG-16, AlexNet, GoogleNet, and ResNet-50 were fine-tuned for brain tumor classification. Results Transfer learning proved highly effective for classifying tumor types and identifying non-tumor images. The performance metrics showed that AlexNet achieved an accuracy of 97.33%, ResNet-50 reached 98.03%, GoogLeNet attained 98.09%, and VGG-16 matched AlexNet with 97.33% accuracy on the test dataset. Among the models, GoogleNet achieved the highest accuracy with fewer parameters, while ResNet-50 demonstrated the highest sensitivity and F1-Score. All models performed exceptionally well, with sensitivity and precision exceeding 95%. Conclusion The research highlights the effectiveness of deep learning techniques, especially CNNs and transfer learning, in classifying brain tumor types using MR images. The study emphasizes that merely increasing model parameters does not guarantee better performance, underscoring the need for fine-tuning models originally trained on ImageNet for medical imaging tasks. These findings contribute to advancing automated medical image analysis, potentially leading to earlier brain tumor diagnosis and improved patient outcomes.