Deep Multi-Scale 3D Convolutional Neural Network (CNN) for MRI Gliomas Brain Tumor Classification.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Radiology
• Neuro-oncology

Background:

• Accurate brain tumor grading is crucial for clinical diagnosis and treatment planning.
• Current methods often require manual analysis, which can be time-consuming and subjective.
• Volumetric 3D MRI offers rich data for automated analysis.

Purpose of the Study:

• To develop an efficient and fully automatic deep learning architecture for brain tumor classification.
• To distinguish between low-grade gliomas (LGG) and high-grade gliomas (HGG) using 3D MRI.
• To evaluate the impact of preprocessing and data augmentation on classification accuracy.

Main Methods:

• Proposed a deep multi-scale 3D convolutional neural network (CNN) architecture.
• Utilized volumetric T1-Gado MRI sequences for classification.
• Implemented intensity normalization, adaptive contrast enhancement, and data augmentation for preprocessing.
• Compared the 3D CNN with a 2D CNN variant.

Main Results:

• The proposed 3D CNN architecture effectively merged local and global contextual information.
• Preprocessing and data augmentation significantly improved classification accuracy.
• Achieved an overall accuracy of 96.49% on the Brats-2018 benchmark dataset.
• Outperformed recent state-of-the-art supervised and unsupervised approaches.

Conclusions:

• The developed deep multi-scale 3D CNN is a promising tool for automated brain tumor grading.
• Effective MRI preprocessing and data augmentation are essential for accurate CNN-based classification.
• The approach offers significant potential for assisting neuroradiologists in clinical diagnosis.