Rubik-Net: Learning Spatial Information via Rotation-Driven Convolutions for Brain Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate brain Magnetic Resonance Image (MRI) segmentation is crucial for diagnosing neurological conditions.
• Challenges include low tissue contrast and partial volume effects, which 2-D convolutional networks struggle to address due to overlooked spatial information between slices.
• Existing 3-D convolutions capture volumetric data but may lead to overfitting with limited data.

Purpose of the Study:

• To introduce a novel convolutional mechanism, Rubik convolution, for capturing multi-dimensional information between MRI slices.
• To propose an efficient 2-D convolutional network, Rubik-Net, leveraging residual connections and bottleneck structures.
• To enhance the accuracy and efficiency of medical image segmentation.

Main Methods:

• Rubik convolution rotates slice axes, allowing 2-D kernels to extract features from multiple axial planes simultaneously.
• Feature maps are rotated back and fused using Max-View-Maps to integrate multidimensional information.
• Rubik-Net incorporates residual connections and bottleneck structures to improve information transmission and reduce parameters.

Main Results:

• Rubik-Net demonstrated promising segmentation accuracy on the iSeg2017, iSeg2019, IBSR, and BrainWeb datasets.
• Achieved state-of-the-art results in 95th percentile Hausdorff distance and average surface distance for cerebrospinal fluid segmentation on the iSeg2019 dataset.
• Experiments confirmed improved accuracy and efficiency in medical image segmentation.

Conclusions:

• Rubik-Net effectively enhances medical image segmentation accuracy and efficiency.
• Rubik convolution offers a versatile approach that can be integrated into existing 2-D convolutional networks.
• The proposed method addresses limitations of traditional 2-D and 3-D convolutions for volumetric MRI data.