3D brain tumor segmentation using an improved V-Net architecture and 3D attention gate

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Neuroimaging

Background:

• Accurate tumor segmentation is vital for cancer diagnosis and treatment planning.
• Manual segmentation is time-consuming and prone to variability.
• Automated methods offer enhanced accuracy, stability, and efficiency.

Purpose of the Study:

• To develop and evaluate an improved 3D V-Net model for automated brain tumor segmentation.
• To segment whole tumors, tumor cores, and enhanced tumors using multimodal MRI data.
• To enhance the performance of the V-Net architecture for 3D medical image analysis.

Main Methods:

• Utilized the BraTS2021 dataset for 3D brain tumor segmentation.
• Applied N4 Bias Field Correction for MRI data preprocessing.
• Implemented an improved 3D V-Net with enhanced encoder (residual and dilated convolutions) and decoder (Attention Gate) components.
• Inputted all four MRI modalities into the 3D improved V-Net.

Main Results:

• Achieved Dice coefficients of 87% for whole tumors, 81.2% for tumor cores, and 74.43% for enhanced tumors.
• Demonstrated the effectiveness of residual and dilated convolutions in the encoder for performance enhancement.
• Showcased the benefit of the Attention Gate in the decoder for improved segmentation accuracy.
• Validated the model's capability in segmenting complex 3D brain tumor structures.

Conclusions:

• The proposed improved 3D V-Net model effectively segments brain tumors in 3D MRI images.
• The architectural enhancements significantly boosted the segmentation performance compared to standard V-Net.
• This automated approach offers a promising tool for clinical applications in neuro-oncology.