A Multi-Scale Liver Tumor Segmentation Method Based on Residual and Hybrid Attention Enhanced Network with Contextual...

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Liver cancer poses a significant global health challenge due to high mortality rates.
• Accurate segmentation of liver and tumors is critical for effective diagnosis and treatment planning.
• Existing methods, including U-Net, face limitations in capturing intricate image details for precise segmentation.

Purpose of the Study:

• To develop an advanced deep learning model, RHEU-Net, for improved liver and tumor segmentation in medical images.
• To enhance the feature extraction capabilities and gradient stability of the U-Net architecture.
• To optimize the fusion of multi-scale and attention-guided features for superior segmentation performance.

Main Methods:

• Proposed RHEU-Net model, an enhanced U-Net architecture.
• Incorporated improved residual modules in the encoder and decoder for better feature extraction.
• Integrated a Hybrid Gated Attention (HGA) module for parallel channel and spatial attention processing.
• Introduced a Multi-Scale Feature Enhancement (MSFE) layer at the bottleneck for richer contextual information.

Main Results:

• RHEU-Net achieved a Dice score of 95.72% for liver segmentation on the LiTS2017 dataset.
• RHEU-Net achieved a Dice score of 70.19% for tumor segmentation on the LiTS2017 dataset.
• The enhanced model demonstrated superior performance in capturing fine image features and details.

Conclusions:

• RHEU-Net effectively improves upon the U-Net architecture for liver cancer image segmentation.
• The model's enhancements in feature extraction and detail preservation show significant potential for clinical applications.
• Accurate segmentation using RHEU-Net can aid in timely detection and diagnosis, improving patient prognosis.