Computed Tomography Radiomics and Machine Learning for Prediction of Histology-Based Hepatic Steatosis Scores.

Area of Science:

• Radiology
• Medical Imaging
• Machine Learning

Background:

• Hepatic steatosis diagnosis often requires invasive liver biopsies.
• Current non-invasive methods like CT evaluation and simple thresholding are insufficient for accurate steatosis assessment.
• There is a need for objective, quantitative, and non-invasive methods to predict steatosis severity.

Purpose of the Study:

• To explore CT radiomics and machine learning for non-invasive prediction of hepatic steatosis severity in macaques.
• To identify a radiomic signature correlating with histological steatosis scores.
• To assess the potential of machine learning models in classifying and quantifying steatosis.

Main Methods:

• Retrospective analysis of CT images from 42 crab-eating macaques with varying degrees of hepatic steatosis.
• Extraction of radiomic features from CT images.
• Application of statistical analyses, feature selection, and machine learning models (k-nearest neighbors) compared against histology-based scores.

Main Results:

• Identified 12 radiomic features correlated with steatosis scores.
• Hierarchical clustering based on radiomics aligned with steatosis severity groups.
• The k-nearest neighbors model achieved high accuracy (AUC ROC = 0.89 ± 0.09) in predicting steatosis, identifying seven key radiomic features.

Conclusions:

• A CT radiomic signature for hepatic steatosis was identified.
• Machine learning and CT radiomics can objectively and non-invasively predict steatosis severity.
• Findings show potential for translation to human liver steatosis assessment due to conserved pathophysiology.