Decoupled pyramid correlation network for liver tumor segmentation from CT images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Automated liver tumor segmentation from CT images is crucial for hepatic abnormality interventions and surgery planning.
• Accurate segmentation is challenging due to variations in tumor size and texture.
• Fully Convolutional Networks (FCNs) have advanced medical image segmentation through pyramid features.

Purpose of the Study:

• To propose a Decoupled Pyramid Correlation Network (DPC-Net) for enhanced liver tumor segmentation.
• To leverage attention mechanisms for improved utilization of low- and high-level features in FCNs.
• To address the challenges of variability in tumor size and inhomogeneous texture.

Main Methods:

• A Pyramid Feature Encoder (PFE) was designed to extract multilevel image features.
• Features were decoupled based on spatial (height, width, depth) and semantic (channel) dimensions.
• Spatial Correlation (SpaCor) and Semantic Correlation (SemCor) attention modules were developed to measure feature correlations.

Main Results:

• DPC-Net achieved a Dice Similarity Coefficient (DSC) of 76.4% and an Average Symmetric Surface Distance (ASSD) of 0.838 mm for liver tumor segmentation.
• The method outperformed state-of-the-art techniques on the MICCAI 2017 LiTS dataset.
• Competitive results were obtained for liver segmentation (DSC: 96.0%, ASSD: 1.636 mm).

Conclusions:

• DPC-Net demonstrates promising performance for segmenting liver and tumors in CT images.
• The proposed SemCor and SpaCor modules effectively model multilevel correlations across semantic and spatial dimensions.
• These lightweight attention modules are extensible to other multilevel methods in an end-to-end fashion.