Three-Dimensional Liver Image Segmentation Using Generative Adversarial Networks Based on Feature Restoration.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer-Aided Diagnosis

Background:

• Accurate liver segmentation in CT images is crucial for liver cancer diagnosis and treatment.
• Challenges include heterogeneous liver tissue, blurred boundaries, and limited labeled 3D data.
• Existing 3D U-Net methods show limitations in segmentation accuracy.

Purpose of the Study:

• To improve the accuracy of 3D liver segmentation from abdominal CT images.
• To address the challenge of insufficient labeled 3D training data.
• To develop a robust semi-supervised algorithm for liver segmentation.

Main Methods:

• An improved 3D U-Net architecture was developed.
• A generative adversarial network (GAN) framework was integrated for semi-supervised learning.
• Deep convolutional neural networks (DCNN) with feature restoration were used for realistic synthetic image generation.

Main Results:

• The proposed semi-supervised 3D liver segmentation method significantly improved performance.
• Achieved a Dice score of 0.9424 on the LiTS-2017 and KiTS19 datasets.
• Outperformed existing segmentation methods.

Conclusions:

• The enhanced semi-supervised 3D U-Net effectively improves liver segmentation accuracy.
• The method offers a promising solution for computer-aided diagnosis of liver cancer.
• The approach addresses data scarcity and image quality issues in medical imaging.