Multi-Scale Convolutional Attention and Structural Re-Parameterized Residual-Based 3D U-Net for Liver and Liver Tumor...

Area of Science:

• Medical Image Analysis
• Artificial Intelligence in Healthcare
• Computational Biology

Background:

• Accurate liver and liver tumor segmentation is vital for clinical decision-making.
• Challenges include complex tumor morphology, small target indistinctness, and similar grayscale values with surrounding organs.

Purpose of the Study:

• To develop an enhanced 3D UNet architecture for improved liver and tumor segmentation.
• To address the limitations of current segmentation methods, especially for small and indistinct targets.

Main Methods:

• Proposed an enhanced 3D UNet architecture named ELANRes-MSCA-UNet.
• Incorporated a structural re-parameterized residual module (ELANRes) and a multi-scale convolutional attention module (MSCA).
• Employed a two-stage segmentation strategy: first segmenting the liver, then segmenting tumors.

Main Results:

• Achieved Dice scores of 97.2% for liver segmentation and 72.9% for tumor segmentation on the LiTS2017 dataset.
• Demonstrated significant performance improvement over state-of-the-art methods, particularly in segmenting small targets.
• The two-stage strategy effectively reduced false positive rates.

Conclusions:

• The ELANRes-MSCA-UNet demonstrates high accuracy and robustness in medical image segmentation.
• The proposed method shows significant potential for enhancing clinical diagnosis and treatment planning for liver diseases.