Towards bi-directional skip connections in encoder-decoder architectures and beyond.

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Computer Vision

Background:

• U-Net architecture is effective for medical image analysis but often computationally expensive.
• Increasing network complexity in U-Net variants hinders clinical application efficiency.

Purpose of the Study:

• To develop an efficient and effective medical image segmentation network.
• To reduce computational cost and improve inference speed in U-Net based models.

Main Methods:

• Proposed backward skip connections integrated with existing forward skip connections in an encoder-decoder framework.
• Introduced Bi-directional O-shape networks (BiO-Nets) and a two-phase Neural Architecture Search (NAS) algorithm (BiX-NAS).
• Evaluated methods on 2D and 3D segmentation tasks across six public medical imaging datasets.

Main Results:

• Bi-directional O-shape networks set new benchmarks on multiple medical imaging segmentation datasets.
• The searched BiX-Net achieved the lowest network complexity.
• BiX-Net outperformed state-of-the-art methods by significant margins.

Conclusions:

• Backward skip connections offer an effective and parameter-efficient enhancement for U-Net architectures.
• BiX-NAS successfully identifies optimal multi-scale bi-directional skip connections for efficient segmentation.
• The proposed methods significantly advance the state-of-the-art in medical image segmentation.