Metabolic-associated fatty liver disease is one of the most prevalent chronic liver conditions worldwide and is closely associated with metabolic syndrome. Affected individuals often present with central obesity, type 2 diabetes mellitus and insulin resistance. The disease spans a continuum from simple hepatic steatosis to more advanced stages that may include inflammation, fibrosis, cirrhosis and hepatocellular carcinoma. Reported prevalence varies substantially, reflecting differences in diagnostic thresholds and methods. Liver biopsy remains the histopathological reference standard but its invasiveness limits routine use, particularly for screening or monitoring. Imaging techniques such as ultrasound and computed tomography are widely used because they are accessible and relatively low cost, though they show limited sensitivity for mild steatosis and reduced accuracy in certain patient groups. More quantitative approaches can improve measurement but are less available. Artificial intelligence has therefore been explored as a means of improving detection and assessment of hepatic steatosis across imaging and pathology data.
Evidence Base and Review Design
A systematic review and meta-analysis evaluated the diagnostic performance of artificial intelligence models for detecting hepatic steatosis, with additional analysis of heterogeneity and clinical applicability. Searches were conducted across major medical and technical databases up to September 2025. Eligible studies applied either deep learning or traditional machine learning models to input images derived from ultrasound, computed tomography, magnetic resonance imaging or pathology.
Must Read: DL Reconstruction for Better and Faster Liver DWI
Reference standards varied and included magnetic resonance imaging–based quantification, histopathological assessment or expert interpretation of ultrasound images. This diversity reflected real-world diagnostic practice but also introduced variability. Included studies were required to report sufficient diagnostic data to allow calculation of sensitivity, specificity and related performance measures. Statistical pooling used random effects modelling to account for between-study variation. Methodological quality was assessed using a structured risk of bias framework covering patient selection, index testing, reference standards and flow and timing. Clinical relevance was examined through likelihood ratios and posttest probability estimates, providing insight into potential rule-in and rule-out capability.
Diagnostic Performance Across Studies
Thirty-six studies met inclusion criteria, with thirty-three studies contributing thirty-six datasets to subgroup analyses. Overall, artificial intelligence models demonstrated strong diagnostic discrimination for hepatic steatosis. Pooled estimates showed high sensitivity and specificity, with area under the curve values indicating excellent separation between steatotic and non-steatotic cases. Clinical applicability analysis suggested that positive results substantially increased the probability of hepatic steatosis, while negative results markedly reduced it, supporting potential use for both confirmation and exclusion. Despite these favourable pooled results, heterogeneity was substantial across studies.
Variability was linked to differences in algorithm type, imaging modality, reference standard, study design and data source. Deep learning approaches generally achieved higher discrimination than traditional machine learning methods. Use of transfer learning was associated with higher sensitivity and stronger exclusion capability. Multicentre studies showed more consistent performance and lower variability compared with single-centre designs, while retrospective studies tended to report higher accuracy than prospective studies, reflecting differences between controlled and real-world settings.
Sources of Variation and Methodological Limitations
Differences in imaging modality and reference standards were major contributors to performance variation. Models based on histopathological inputs achieved the highest accuracy but face practical constraints related to invasiveness, cost and patient acceptance. Magnetic resonance imaging–based approaches provided precise, non-invasive quantification but were limited by availability and confounding factors such as iron overload or concurrent pathology. Ultrasound-based models were the most accessible and economical, though performance was affected by operator dependency, reduced sensitivity for mild steatosis and technical limitations in obese patients. Computed tomography approaches showed strong diagnostic accuracy but were constrained by radiation exposure and the influence of contrast agents.
Methodological assessment identified frequent risk of bias, particularly in patient selection, where limited representativeness and incomplete reporting were common. Additional concerns included inconsistent image quality control, non-standardised model training and validation and unclear blinding of reference standards. Reporting of patient flow and timing was often insufficient, complicating interpretation of results. Public datasets were associated with high reported accuracy but may not reflect clinical diversity, while prospective studies better captured real-world conditions at the expense of lower measured performance.
Artificial intelligence models demonstrate strong potential for detecting hepatic steatosis across multiple imaging and pathology inputs, with pooled evidence indicating high diagnostic accuracy. Performance varied widely across studies, influenced by imaging modality, reference standard, algorithm design and study methodology. Methodological bias and limited prospective validation remain key challenges. Translation into routine practice depends on improving study quality, expanding data diversity and integrating systems into existing clinical workflows while addressing cost, privacy and interoperability constraints.
Source: Journal of Medical Internet Research
Image Credit: iStock
