GT-Net: global transformer network for multiclass brain tumor classification using MR images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Multiclass classification of brain tumors using magnetic resonance (MR) images presents significant challenges due to high inter-class similarities.
• Conventional Convolutional Neural Networks (CNNs) often struggle to identify small, intricate lesion patterns crucial for accurate tumor classification.

Purpose of the Study:

• To propose a novel Global Transformer Network (GT-Net) designed to enhance multiclass brain tumor classification from MR images.
• To address the limitations of existing CNN architectures in capturing subtle tumor lesion details.

Main Methods:

• Introduction of GT-Net, featuring a Global Transformer Module (GTM) integrated with a backbone network.
• Development of a Generalized Self-Attention Block (GSB) within the GTM to capture both spatial and channel feature inter-dependencies.
• Utilization of multiple GSB heads in the GTM to leverage global feature dependencies for improved analysis.

Main Results:

• The proposed GT-Net demonstrated effectiveness in multiclass brain tumor classification across various backbone networks.
• The Global Transformer Module (GTM) proved adept at extracting detailed tumor lesion information while filtering out less relevant data.
• Comparative analysis confirmed the superiority of GT-Net over current state-of-the-art methods on a benchmark dataset.

Conclusions:

• GT-Net offers a significant advancement in brain tumor classification by effectively addressing the challenge of subtle lesion pattern detection.
• The proposed GTM and GSB architecture provide a robust framework for leveraging global feature dependencies in medical image analysis.
• The findings highlight the potential of transformer-based networks in improving diagnostic accuracy for complex medical imaging tasks.