Liver tissue segmentation in multiphase CT scans using cascaded convolutional neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of liver tissues in CT scans is crucial for diagnosing and monitoring hepatocellular carcinoma (HCC).
• Distinguishing between healthy parenchyma, active tumor, and necrotic regions within HCC is clinically significant.

Purpose of the Study:

• To develop and evaluate a deep learning approach for automatic segmentation of liver tissues, including HCC and its necrotic components, on multiphase CT images.
• To compare different strategies for incorporating multiphase information into a cascaded U-Net architecture.

Main Methods:

• A cascaded convolutional neural network based on the U-Net architecture was developed.
• Two methods for handling multiphase CT data were compared: input layer concatenation and independent phase processing with output merging.
• Specialized networks within the cascade were designed for segmenting specific tissue types.

Main Results:

• The proposed cascaded multiphase method achieved performance comparable to state-of-the-art MR segmentation techniques and surpassed previous CT segmentation methods.
• Cascaded specialized networks demonstrated higher prediction accuracy than a single network performing all segmentation tasks.
• Multiphase information significantly improved the segmentation of cancerous tissues, particularly the active versus necrotic parts of the tumor.

Conclusions:

• The developed cascaded multiphase architecture shows strong potential for accurate automatic liver tissue segmentation in CT imaging.
• This method enables reliable estimation of the tumor necrosis rate, a key biomarker for predicting clinical outcomes in HCC patients.