Detecting liver cirrhosis in computed tomography scans using clinically-inspired and radiomic features

Area of Science:

• Medical Imaging
• Radiology
• Computational Pathology

Background:

• Hepatic cirrhosis is a leading cause of mortality, representing the final stage of chronic liver disease.
• Early detection of asymptomatic cirrhosis is crucial to prevent hospitalization and fatal outcomes.
• Current methods for assessing liver state from CT scans are subjective and lack reproducibility.

Purpose of the Study:

• To propose an end-to-end, reproducible approach for detecting hepatic cirrhosis using abdominal computed tomography (CT) scans.
• To integrate clinically-inspired features with radiomic features for improved diagnostic accuracy.
• To enhance the interpretability and efficiency of cirrhosis detection models.

Main Methods:

• Development of an automated system for cirrhosis detection from CT images.
• Extraction of clinically-inspired features reflecting patient characteristics.
• Extraction of radiomic features from the liver and its rectified contour.
• Integration and selection of discriminative features using supervised learning models.
• Validation on two heterogeneous clinical datasets (241 and 32 patients).

Main Results:

• Extracting radiomic features from the liver's rectified contour significantly improved classification performance.
• Incorporating clinically-inspired image features enhanced model performance and identified key diagnostic indicators.
• Feature selection dramatically reduced the number of features (280×), leading to Pareto-optimal models.
• The proposed approach demonstrated enhanced feature-level interpretability.

Conclusions:

• The developed method offers a reproducible and accurate approach for early hepatic cirrhosis detection from CT scans.
• Combining radiomic and clinically-inspired features is pivotal for improving diagnostic capabilities.
• The approach enhances model interpretability and efficiency, aiding clinical decision-making in managing chronic liver disease.