Lean Unet: a compact model for image segmentation.

Area of Science:

• Medical image analysis
• Deep learning for radiology
• Computer-assisted diagnosis

Background:

• Unet architectures are standard for semantic image segmentation in computer-assisted radiology.
• Current Unet models have large memory footprints and high inference latency due to iterative downsampling and channel expansion.
• Existing channel pruning methods require extensive optimization and may lack generalizability.

Purpose of the Study:

• To investigate Unet pruning strategies and identify key architectural factors for efficient segmentation.
• To propose a novel, lean Unet (LUnet) architecture with a compact, flat hierarchy.
• To reduce memory footprint and inference latency without compromising segmentation accuracy.

Main Methods:

• Investigated Unet pruning to hypothesize that network structure, not channel selection, is critical.
• Proposed LUnet, a lean Unet architecture with a constant channel count and flat hierarchy.
• Evaluated LUnet on public MRI and internal CT datasets, comparing against conventional Unet and pruned networks (STAMP+).

Main Results:

• LUnet achieved performance comparable to conventional Unet and data-adaptively pruned networks.
• LUnet demonstrated over 30 times fewer parameters than conventional counterparts.
• Simple channel elimination strategies performed similarly or better than advanced pruning methods like STAMP+.

Conclusions:

• The proposed lean Unet (LUnet) offers superior performance with significantly fewer parameters compared to standard Unet.
• LUnet's constant channel count and efficient skip connections reduce the need for increased bottleneck channels.
• This architecture enhances efficiency for medical image segmentation tasks.