An improved residual U-Net with morphological-based loss function for automatic liver segmentation in computed...

Area of Science:

• Medical Imaging
• Computer Vision
• Artificial Intelligence

Background:

• Accurate liver segmentation in CT scans is crucial for diagnosis and treatment planning.
• Existing segmentation methods face challenges with pixel imbalance and generalization.
• U-Net and ResU-Net architectures are foundational but can be further optimized.

Purpose of the Study:

• To develop an improved ResU-Net framework for enhanced automatic liver CT segmentation.
• To increase segmentation accuracy and model generalization capabilities.
• To validate the proposed method on public liver CT datasets.

Main Methods:

• Implemented a residual module to replace standard convolution layers in U-Net for faster convergence.
• Introduced an inverse Dice loss function to address pixel imbalance, weighted by morphological methods.
• Utilized random affine transformation and elastic deformation for data augmentation to improve generalization.

Main Results:

• Achieved high segmentation accuracies: DICE global (94.28%), DICE per case (94.24 ± 2.07), VOE (10.83 ± 3.70), and RVD (-0.25 ± 2.74).
• Demonstrated significantly improved performance compared to standard U-Net and ResU-Net.
• Showcased faster convergence speed, stronger generalization ability, and robustness on 2D and 3D data, even with low-contrast organs.

Conclusions:

• The proposed improved ResU-Net framework offers superior liver CT segmentation performance.
• The novel loss function and augmentation strategies enhance accuracy and model robustness.
• The method is scalable and effective for diverse clinical imaging scenarios.