SPA-UNet: A liver tumor segmentation network based on fused multi-scale features

Area of Science:

• Medical image analysis
• Artificial intelligence in oncology
• Deep learning for medical segmentation

Background:

• Accurate liver tumor segmentation is crucial for liver cancer diagnosis and treatment.
• Traditional U-shaped convolutional neural networks (UNets) face challenges in precise boundary delineation and small tumor detection, impacting segmentation accuracy.

Purpose of the Study:

• To enhance the accuracy of liver tumor segmentation using a novel deep learning architecture.
• To address limitations in current UNet models for segmenting liver tumors, particularly small ones and their boundaries.

Main Methods:

• Proposes Space Pyramid Attention (SPA)-UNet, an encoder-decoder network.
• Incorporates Spatial Pyramid Convolution Block (SPCB) for multi-scale feature extraction and Spatial Pyramid Pooling Block (SPPB) for downsampling.
• Utilizes an Upsample module and Residual Attention Block (RA-Block) in the decoder for precise localization.

Main Results:

• SPA-UNet achieved a 1.0% and 2.0% improvement in intersection over union (IoU) for liver and tumors, respectively, compared to traditional UNet.
• Demonstrated increased recall rates by 1.2% for liver and 1.8% for tumors.
• Experimental results on a liver tumor segmentation dataset validated the model's superior performance.

Conclusions:

• SPA-UNet significantly improves liver tumor segmentation accuracy and detection capabilities.
• The proposed model offers a dependable foundation for advancing liver cancer diagnosis and treatment planning.
• Enhanced segmentation performance contributes to more effective clinical decision-making in oncology.