Performance of convolutional neural networks for the classification of brain tumors using magnetic resonance imaging

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Brain tumor classification is complex due to tumor heterogeneity.
• Deep learning shows promise for accurate tumor detection and classification.
• A standardized comparison of deep learning techniques for brain tumor classification is lacking.

Purpose of the Study:

• To analyze and compare the performance of various convolutional neural networks (CNNs) for brain tumor classification.
• To evaluate different deep learning architectures including custom CNNs, VGG, ResNet, EfficientNet, and ConvNeXt.
• To assess the impact of training strategies like transfer learning and fine-tuning on classification accuracy.

Main Methods:

• Utilized two magnetic resonance imaging (MRI) datasets comprising over 3000 images of gliomas, meningiomas, pituitary tumors, and healthy tissues.
• Implemented and compared a custom CNN, VGG, ResNet, EfficientNet, and ConvNeXt architectures.
• Evaluated models using training from scratch, data augmentation, transfer learning, and fine-tuning, optimizing hyperparameters on training and validation sets.

Main Results:

• Several CNNs achieved high accuracy, with the best model reaching 98.7%.
• Average precision varied by tumor type: gliomas (94.3%), meningiomas (93.8%), pituitary tumors (97.9%), and non-tumor images (95.3%).
• MobileNet and EfficientNet demonstrated superior performance regarding accuracy and computational complexity, with fast training times and high image throughput.

Conclusions:

• Deep learning models, particularly efficient architectures like MobileNet and EfficientNet, are highly effective for brain tumor classification using MRI data.
• Transfer learning combined with fine-tuning is the most effective strategy for maximizing classification accuracy.
• While data augmentation did not consistently improve accuracy, efficient models offer a balance of high performance and computational feasibility.