Brain tumor grade classification using the ConvNext architecture.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Brain tumor grading is crucial for treatment planning.
• Traditional methods like biopsy are invasive and can be inaccurate.
• Deep learning offers non-invasive, accurate brain tumor diagnosis, but faces data scarcity challenges.

Purpose of the Study:

• To develop a non-invasive brain tumor grade classification technique using a modern convolutional neural network (CNN).
• To leverage the ConvNext architecture for feature extraction from magnetic resonance imaging (MRI) data.
• To address data scarcity in medical imaging through transfer learning and advanced CNN designs.

Main Methods:

• Utilized the ConvNext architecture for feature extraction from MRI data.
• Employed transfer learning with a pre-trained ConvNext model.
• Fed extracted features into a fully connected neural network for classification.
• Input three MRI sequences as channels into the CNN.

Main Results:

• Achieved state-of-the-art performance on the BraTS 2019 dataset.
• Obtained a maximum classification accuracy of 99.5%.
• Demonstrated superior performance using three MRI sequences as input channels.

Conclusions:

• The proposed method using ConvNext CNN is highly effective for brain tumor grade classification.
• Modern CNNs, like ConvNext, possess strong inductive biases beneficial for image data compared to vision transformers.
• This deep learning approach provides a promising non-invasive tool for brain tumor diagnosis.