KiU-Net: Overcomplete Convolutional Architectures for Biomedical Image and Volumetric Segmentation.

Area of Science:

• Medical Image Analysis
• Computer Vision
• Deep Learning

Background:

• U-Net and its variants are standard for medical image segmentation but struggle with small structures and precise boundaries.
• This limitation stems from the increasing receptive field in encoder-decoder networks, which dilutes low-level feature information crucial for fine details.

Purpose of the Study:

• To address the limitations of traditional encoder-decoder segmentation networks.
• To propose a novel architecture, KiU-Net, for improved medical image segmentation, particularly for small structures and boundaries.

Main Methods:

• Introduced KiU-Net, a hybrid architecture combining an overcomplete convolutional network (Kite-Net) for fine details and a U-Net for high-level features.
• Developed KiU-Net 3D for volumetric segmentation, utilizing 3D convolutions.
• Investigated extensions using residual and dense blocks for further enhancements.

Main Results:

• KiU-Net demonstrated superior performance in segmenting small structures and precise boundaries across five diverse datasets.
• The proposed architecture achieved competitive results with fewer parameters and faster convergence compared to existing methods.
• Extensions with residual and dense blocks further boosted performance.

Conclusions:

• KiU-Net offers a significant advancement in medical image segmentation by effectively capturing fine details and high-level features.
• The architecture provides a more efficient and accurate solution for segmenting challenging structures in various medical imaging modalities.