RDCTrans U-Net: A Hybrid Variable Architecture for Liver CT Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate medical image segmentation is crucial for diagnosing diseases and planning treatments.
• U-Net based models are commonly employed for liver tumor segmentation, but challenges remain due to tumor variability.
• Existing methods struggle with the diverse shapes and sizes of liver tumors.

Purpose of the Study:

• To propose a novel U-Net based hybrid model, RDCTrans U-Net, for enhanced liver tumor segmentation in computed tomography (CT) examinations.
• To improve the accuracy and efficiency of liver tumor segmentation by integrating advanced architectural components.
• To address the limitations of existing models in handling the variability of tumor shapes and sizes.

Main Methods:

• Developed a hybrid variable structure U-Net model (RDCTrans U-Net) for liver tumor segmentation.
• Incorporated a backbone network using ResNeXt50 and dilated convolutions to enhance feature extraction efficiency and perceptual field.
• Integrated Transformer into the down-sampling layers to improve global image understanding and segmentation accuracy.

Main Results:

• The RDCTrans U-Net model achieved 89.22% mean Intersection over Union (mIoU) and 98.91% accuracy for liver and tumor segmentation on the LiTS dataset.
• The model obtained Dice coefficients of 93.38% for liver and 89.87% for tumor segmentation.
• Demonstrated state-of-the-art (SOTA) performance compared to original U-Net and variants incorporating dense connections, attention mechanisms, or Transformers.

Conclusions:

• The proposed RDCTrans U-Net model significantly enhances liver tumor segmentation accuracy in CT images.
• The hybrid architecture effectively addresses challenges posed by variable tumor shapes and sizes.
• This approach represents a significant advancement in medical image analysis for liver tumor detection and treatment planning.