Liver segmentation network based on detail enhancement and multi-scale feature fusion

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Abdominal CT image segmentation, particularly for the liver, is difficult due to low contrast and similar organ morphologies.
• Varying imaging planes (sagittal, coronal, transverse) further complicate liver segmentation by increasing morphological diversity.

Purpose of the Study:

• To develop an advanced deep learning model for precise liver segmentation in abdominal CT images.
• To enhance the learning of subtle liver features and multi-scale contextual information for improved segmentation accuracy.

Main Methods:

• Proposed a novel Detail Enhanced Convolution (DE Conv) to improve the extraction of liver-specific features.
• Introduced a Multi-Scale Feature Fusion (MSFF) module within skip connections to capture global and multi-scale liver characteristics.
• Developed the Detail Enhancement and Multi-Scale Feature Fusion Network (DEMF-Net) integrating these novel components.

Main Results:

• DEMF-Net demonstrated significant improvements in liver segmentation performance on the LiTS17 dataset.
• The model achieved enhanced accuracy across multiple standard evaluation metrics for segmentation tasks.
• The DE Conv and MSFF modules effectively addressed challenges related to low contrast and morphological variations.

Conclusions:

• The proposed DEMF-Net effectively overcomes the limitations of traditional methods for liver segmentation in abdominal CT.
• The integration of detail enhancement and multi-scale feature fusion leads to highly accurate and robust liver segmentation.
• DEMF-Net represents a significant advancement in automated medical image segmentation for liver analysis.