Segmenting brain glioblastoma using dense-attentive 3D DAF²

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Oncology
• Neuroimaging

Background:

• Accurate brain glioblastoma segmentation is crucial for patient diagnosis and treatment monitoring.
• Multimodal MRI segmentation is challenging due to varying intensity profiles and tumor biological properties.
• Convolutional Neural Networks (CNNs) show promise for glioblastoma segmentation.

Purpose of the Study:

• To introduce an innovative methodology for brain glioblastoma segmentation using a Dense-Attention 3D U-Net.
• To enhance tumor boundary discernment and address class imbalance in medical image analysis.
• To improve the efficiency and performance of glioblastoma segmentation models.

Main Methods:

• Developed a Dense-Attention 3D U-Net combined with a fusion strategy and focal Tversky loss function.
• Fused information from multiple resolution segmentation maps for improved boundary detection.
• Incorporated Recursive Convolution Block 2 for efficient feature utilization and rapid convergence.

Main Results:

• Achieved an average Dice Similarity Coefficient (DSC) of 82.4% and an average Hausdorff Distance 95th percentile (HD95) of 10.426.
• Demonstrated comparable performance to other methods with increased efficiency.
• Showed consistent performance improvement over baseline models (U-Net, Attention U-Net, V-Net, Res U-Net).

Conclusions:

• The proposed Dense-Attention 3D U-Net with fusion and focal Tversky loss offers an effective approach for brain glioblastoma segmentation.
• The methodology enhances segmentation accuracy and efficiency in multimodal MRI.
• Results indicate significant potential for clinical application in neuro-oncology.