Liver tumor segmentation using 2.5D UV-Net with multi-scale convolution.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Liver tumor segmentation often uses 2D or 3D U-Net architectures.
• 2D networks miss inter-slice context, while 3D networks demand substantial GPU memory.
• 2.5D networks offer a balance, but canonical designs have limitations.

Purpose of the Study:

• To introduce UV-Net, a novel 2.5D network for improved liver tumor segmentation.
• To enhance the encoding of inter-slice information and multi-scale feature extraction.
• To achieve superior segmentation accuracy, particularly for small liver tumors.

Main Methods:

• Developed UV-Net, a 2.5D network utilizing 3D convolutions for inter-layer information encoding.
• Incorporated a multi-scale convolution structure for efficient feature extraction.
• Employed a 2D deconvolution approach for high-resolution result reconstruction.
• Utilized a preprocessing method involving the removal of mean energy.

Main Results:

• UV-Net significantly outperforms existing liver tumor segmentation methods.
• Demonstrated improved segmentation accuracy for small objects within the liver.
• The multi-scale convolution structure enhanced network capacity and efficiency.

Conclusions:

• UV-Net offers an effective 2.5D approach for liver tumor segmentation.
• The proposed architecture balances computational cost and contextual information effectively.
• UV-Net shows promise for clinical applications requiring precise liver tumor delineation.