Performance improvement of weakly supervised fully convolutional networks by skip connections for brain structure...

Area of Science:

• Medical Image Analysis
• Neurosurgery Planning
• Deep Learning Architectures

Background:

• Accurate brain structure segmentation in MRI is crucial for neurosurgery.
• Fully convolutional networks (FCNs) are effective for image segmentation.
• Labeling medical images for training FCNs is labor-intensive, necessitating methods that perform well with sparse annotations.

Purpose of the Study:

• To investigate the effectiveness of different skip connection architectures within FCNs for brain structure segmentation using sparsely annotated MRI data.
• To identify which types of skip connections yield optimal performance in FCNs trained with limited annotations.

Main Methods:

• Developed and tested four 2D FCN architectures: U-Net, U-Net++, Full-Resolution Residual Network (FRRN), and a hybrid model.
• All architectures incorporated different skip connection strategies (horizontal, vertical, or combined).
• Models were trained and evaluated under identical conditions using sparsely annotated brain MR images.

Main Results:

• Horizontal skip connections (U-Net) excelled at segmenting larger structures.
• Vertical skip connections (FRRN) improved segmentation of smaller structures like blood vessels.
• The hybrid architecture, combining both horizontal and vertical skip connections, demonstrated superior overall segmentation performance.
• Ablation studies confirmed vertical up paths were particularly effective for small object segmentation.

Conclusions:

• Skip connection design significantly impacts FCN performance in brain MRI segmentation with sparse annotations.
• Hybrid architectures integrating both horizontal and vertical skip connections offer the most effective approach for multiclass brain segmentation.
• The findings provide valuable insights for developing efficient deep learning models for neuroimaging analysis with reduced annotation effort.