A lightweight 3D UNet model for glioma grading

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Glioma, a fatal brain cancer, is classified into low-grade glioma (LGG) and high-grade glioma (HGG).
• Accurate grading of glioma is critical for timely diagnosis and treatment planning.
• Current deep learning algorithms for glioma grading face challenges with numerous parameters, complex computations, and slow processing speeds.

Purpose of the Study:

• To develop a lightweight and efficient deep learning framework for automatic glioma grading.
• To improve classification accuracy compared to existing glioma grading methods.
• To provide a non-invasive diagnostic tool for clinical decision-making in glioma treatment.

Main Methods:

• A novel lightweight 3D UNet deep learning framework was proposed, incorporating depthwise separable convolutions to reduce network parameters.
• A space and channel compression & excitation module was employed to enhance feature maps and model performance.
• Tumor areas were marked using a cube bounding box method, demonstrating comparable performance to manual segmentation.

Main Results:

• The proposed model achieved an accuracy of 89.29% on test datasets.
• Experiments were conducted on MRI images (T1w, T2w, FLAIR, CET1w) from 560 glioma patients.
• The cube bounding box method for tumor marking showed no significant difference in model performance compared to manual ground truth.

Conclusions:

• The developed lightweight 3D UNet framework offers an effective and efficient approach for automatic glioma grading (LGG/HGG).
• This non-invasive method can provide valuable diagnostic suggestions, potentially accelerating clinical treatment decisions.
• The model's efficiency and accuracy contribute to advancing AI applications in neuro-oncology.