Deep Learning-Based Automatic Brain Tumor MRI Classification System: A Comparative Study of Three Transfer Learning Architectures

This article presents a study on a deep learning-based automatic classification system for brain tumor MRI images, comparing the performance of three transfer learning architectures: VGG16, ResNet50, and EfficientNetB0. Using a public dataset, ResNet50 achieved a classification accuracy of 98.9%. The study aims to assist doctors in making fast and accurate diagnoses, providing practical references for AI applications in medical imaging.

Early diagnosis of brain tumors has a decisive impact on patients' survival rates and quality of life. Traditional MRI image diagnosis relies on radiologists' experience, which is time-consuming, labor-intensive, and subject to subjective differences. Deep learning technology has promoted the development of computer-aided diagnosis. This study focuses on automatic classification of brain tumor MRI to assist doctors in decision-making, providing valuable references for AI applications in medical imaging.

Dataset: Adopted the public Brain Tumor MRI Dataset, containing 7023 images divided into four categories: glioma, meningioma, pituitary tumor, and no tumor. Transfer Learning Architectures: Compared three pre-trained CNN architectures:

• VGG16: Fine-tuned at layer 15, added global average pooling, fully connected layers, and Dropout;
• ResNet50: Fine-tuned at layer 143, used residual connections to solve gradient vanishing;
• EfficientNetB0: Fine-tuned at layer 150, balanced model parameters and accuracy through compound scaling. Preprocessing and Augmentation: Uniform image size and pixel normalization; data augmentation included random rotation, brightness adjustment, and histogram equalization (experimental).

Accuracy comparison of three models:

Model	Accuracy
VGG16	97.3%
VGG16 + Histogram Equalization	97.6%
ResNet50	98.9%
EfficientNetB0	93.4%
Key Findings: ResNet50 performed best, with perfect classification for no tumor and pituitary tumor cases; histogram equalization had limited improvement on VGG16; EfficientNetB0 had difficulty distinguishing between glioma and meningioma.

Application Value: Provided a second opinion for radiologists, assisted primary doctors in under-resourced medical areas with screening, and improved diagnostic efficiency and consistency. Future Directions: Extend tumor segmentation functionality, introduce explainable AI (Grad-CAM), conduct multi-center clinical validation, develop web-based diagnostic tools, and integrate with hospital PACS systems.

This study verified the potential of transfer learning in medical image analysis, with ResNet50 performing excellently in brain tumor classification tasks. The study revealed fine-grained classification differences among different architectures, providing references for subsequent research. It also offers a complete practical case for medical AI developers, and AI-assisted diagnostic tools will play an important role in future medical practice.