Deep learning and level set approach for liver and tumor segmentation from CT scans

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare
• Computational Anatomy

Background:

• Manual segmentation of liver and tumors in CT scans is time-consuming and prone to variability.
• Accurate segmentation is crucial for effective hepatic surgical planning, treatment, and follow-up.
• Automated methods are highly desirable to improve efficiency and consistency.

Purpose of the Study:

• To develop an automated method for segmenting liver organs and tumors from computed tomography (CT) scans.
• To improve the accuracy and efficiency of liver and tumor segmentation for clinical applications.
• To provide a tool that assists in hepatic surgical planning and patient follow-up.

Main Methods:

• A novel method combining a fully convolutional neural (FCN) network with a region-based level set function was developed.
• The FCN was trained to predict coarse segmentations of the liver and tumors.
• A localized region-based level set function refined these predictions for accurate final segmentation.

Main Results:

• The method achieved high Dice scores on public datasets: 95.2% for liver and 76.1% for tumors on IRCAD, and 95.6% for liver and 70% for tumors on LiTS.
• Validation on heterogeneous datasets demonstrated the method's robustness and generalizability.
• The automated segmentation showed significant accuracy for both liver and tumor structures.

Conclusions:

• The proposed automated method effectively segments liver and tumors in diverse CT scans, proving its generalization capability.
• This approach shows promise as a valuable tool for the routine clinical analysis of liver and its tumors.
• The developed technique can enhance the precision of surgical planning and follow-up assessments.