Automated CT LI-RADS v2018 scoring of liver observations using machine learning: A multivendor, multicentre...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Hepatocellular Carcinoma (HCC) Detection

Background:

• Computed tomography (CT)/magnetic resonance imaging (MRI) Liver Imaging Reporting and Data System (LI-RADS) v2018 major feature assessment shows significant inter-reader variability.
• This variability can decrease the diagnostic accuracy for hepatocellular carcinoma (HCC).
• A machine learning (ML) algorithm was developed to aid in assessing CT LI-RADS features and liver lesion categorization.

Purpose of the Study:

• To assess the performance and added value of an ML-based algorithm in evaluating CT LI-RADS major features and categorizing liver observations.
• To compare the ML algorithm's performance against qualitative assessment by radiologists.
• To evaluate the sequential use of the ML algorithm and radiologists in a triage and add-on scenario for LR-3/4 lesions.

Main Methods:

• Retrospective analysis of multiphase contrast-enhanced CT scans from high-risk patients with pathologically proven liver lesions (n=429 lesions from 318 patients).
• An ML algorithm was trained to identify key LI-RADS features (non-rim arterial phase hyperenhancement, washout, enhancing capsule).
• Comparison of ML-assisted assessment with two independent readers and a reference standard of three senior readers.

Main Results:

• In the test set, CT LI-RADS v2018 assessment achieved 67% sensitivity and 91% specificity for LR-5, with 70.1% accurate categorization.
• Utilizing the ML algorithm in a triage scenario significantly improved overall performance for LR-5, increasing sensitivity and specificity for independent readers.
• The ML algorithm demonstrated potential in improving lesion categorization accuracy, particularly when used sequentially with radiologists.

Conclusions:

• Quantitative assessment of CT LI-RADS v2018 major features using an ML algorithm is feasible and performs well in diagnosing LR-5 observations.
• The ML algorithm's combination with radiologists' visual analysis shows high performance in high-risk HCC patients.
• AI-enriched diagnostic pathways can standardize and improve liver lesion analysis, especially in non-expert centers, aiding clinical decision-making.