Automatic liver segmentation by integrating fully convolutional networks into active contour models.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate 3D liver segmentation in CT images is crucial for diagnosing and treating liver diseases.
• Fully convolutional networks (FCNs) show promise for medical image segmentation but struggle with precise boundary localization.
• Existing FCNs often lack the ability to enforce local label consistency and boundary smoothness.

Purpose of the Study:

• To develop an automatic and accurate 3D liver segmentation method for CT images, particularly for livers affected by severe diseases.
• To address the limitations of FCNs in accurately segmenting liver boundaries.
• To improve the robustness and reliability of automated liver segmentation in clinical settings.

Main Methods:

• A novel framework integrating FCN predictions with active contour models (ACM) was developed.
• A single network architecture generates pixel label maps with spatial and boundary information.
• An adaptive external constraint force for ACM was designed, leveraging FCN outputs to guide contour evolution.

Main Results:

• The integrated ACM model achieved high mean Dice coefficients (DICE) of 95.8% on clinical data.
• The model demonstrated strong performance on independent datasets (SLIVER07: 96.2%, LiTS: 94.3%) without fine-tuning.
• Significant improvements in surface distance and DICE values were observed compared to FCN alone, with flexible initialization.

Conclusions:

• The proposed model enhances 3D liver segmentation accuracy and boundary definition on CT scans, outperforming FCNs alone.
• The framework effectively handles variations across different datasets due to its inherent deformable nature.
• The method offers a flexible platform for integrating advanced FCN architectures to further boost segmentation performance.