Brain tumor segmentation using neuro-technology enabled intelligence-cascaded U-Net model.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Brain tumors pose significant health risks, necessitating accurate segmentation for clinical identification and treatment.
• Automated brain tumor segmentation (BTS) faces challenges due to tumor variability in size, shape, and location.
• Existing deep learning models like U-Net struggle with limited receptive fields, spatial information loss, and inadequate context.

Purpose of the Study:

• To propose a novel model, Intelligence Cascade U-Net (ICU-Net), for enhanced brain tumor segmentation.
• To address limitations in current segmentation models by incorporating dynamic convolution and non-local attention mechanisms.
• To improve the reconstruction of detailed spatial and contextual information for more accurate BTS.

Main Methods:

• Developed a two-stage 3D U-Net cascade architecture (ICU-Net) with dynamic convolutions and a non-local attention mechanism.
• Applied Expectation-Maximization to lateral connections for improved contextual data utilization.
• Utilized dynamic convolutions with local adaptive capabilities to enhance local feature capture.

Main Results:

• ICU-Net demonstrated high performance in brain tumor segmentation tasks.
• Achieved Dice scores of 0.897/0.903 for tumor core, 0.826/0.828 for complete tumor, and 0.781/0.786 for enhanced tumor on BraTS 2019/2020 validation sets.
• Outperformed other conventional methods in extensive testing.

Conclusions:

• The proposed ICU-Net model significantly advances automated brain tumor segmentation.
• The integration of dynamic convolution and attention mechanisms effectively addresses spatial and contextual information challenges.
• ICU-Net shows strong potential for clinical application in brain tumor diagnosis and treatment planning.