Chronic liver disease often remains clinically silent until advanced stages, making earlier risk assessment difficult. Routine ultrasound, computed tomography and magnetic resonance imaging already support non-invasive evaluation, and artificial intelligence adds new ways to extract quantitative and textural information from these studies. A narrative review published in the British Journal of Radiology describes how these methods are being used to strengthen prognostic modelling for outcomes including survival, decompensation and hepatocellular carcinoma.
Imaging Modalities and AI Integration
Ultrasound remains widely accessible across care settings, from primary care to intensive care environments. Its integration into existing hepatology pathways supports large-scale data collection for artificial intelligence model development. Algorithms applied to ultrasound data perform classification and segmentation tasks, combining imaging and clinical parameters in prognostic models. However, variability between operators, equipment and acquisition techniques introduces inconsistency, particularly in radiomic feature extraction.
Must Read: DL Reconstruction for Better and Faster Liver DWI
Computed tomography provides extensive opportunistic data, particularly in acutely unwell patients and in scans acquired for non-liver indications. Artificial intelligence applications in computed tomography predominantly focus on automated segmentation, enabling extraction of liver and spleen volumes, body composition metrics and radiomic features. Classification approaches are also used, although traditional statistical models remain common. The availability of larger datasets compared with magnetic resonance imaging supports model development.
Magnetic resonance imaging offers a broad range of quantitative data, including volumetric and functional parameters. Artificial intelligence methods applied to magnetic resonance imaging include segmentation and classification, with some models relying solely on imaging data. Despite its potential, longer acquisition times and variability across scanners complicate standardisation, limiting widespread integration into routine clinical pathways.
Prognostic Endpoints and Model Performance
Artificial intelligence-enhanced imaging models address multiple clinically relevant endpoints in chronic liver disease. Prediction of mortality and transplant-free survival incorporates features derived from ultrasound, computed tomography and magnetic resonance imaging. Ultrasound-based parameters such as hepatic echogenicity, contour nodularity and organ dimensions demonstrate predictive value over extended follow-up periods. Computed tomography models integrate body composition metrics, including muscle and adipose tissue indices, as well as radiomic features, to estimate long-term outcomes. Volumetric measurements of the liver and spleen contribute additional prognostic information across both short- and long-term timeframes.
Magnetic resonance imaging parameters, including liver-to-spleen volumetric ratios and enhancement characteristics, associate with survival outcomes over extended periods. These models reflect the capacity of artificial intelligence to combine structural and functional imaging data into prognostic frameworks.
Prediction of clinical decompensation and gastrointestinal bleeding similarly relies on imaging-derived parameters. Ultrasound measurements of spleen size and elastography metrics, combined with physiological variables, identify patients at risk of acute deterioration. Computed tomography models integrate volumetric and radiomic data to predict decompensation across varying time horizons. Magnetic resonance imaging contributes additional markers, including volumetric ratios and textural features, associated with progression to advanced disease.
Models also extend to prediction of hepatocellular carcinoma development. Ultrasound-derived features and computed tomography volumetry demonstrate associations with long-term cancer risk, while combined imaging and radiomic approaches support short-term prediction. Additional applications include prediction of post-treatment outcomes, such as mortality following transjugular intrahepatic portosystemic shunt procedures, post-transplant complications and recurrence following tumour resection. These applications illustrate the breadth of prognostic endpoints addressed through imaging-based artificial intelligence.
Barriers to Clinical Implementation
Several limitations constrain translation of imaging-based artificial intelligence models into routine practice. Model development frequently relies on relatively small datasets, limiting generalisability. Although techniques such as data augmentation and transfer learning support model training, larger and more diverse datasets remain necessary, particularly for outcomes with low event rates. Use of non-medical pretraining datasets introduces further uncertainty regarding model robustness.
Radiomics-based approaches present additional challenges due to the large number of variables evaluated relative to available outcome events. This imbalance increases the risk of overfitting and reduces reliability. Reporting standards vary widely, with incomplete descriptions of model architecture, training processes and validation methods limiting reproducibility. The lack of transparency contributes to the perception of artificial intelligence models as opaque systems, hindering clinical trust.
Prognostic modelling itself introduces further constraints. Many models originate from single-centre datasets or rely on data from a single imaging device, reducing external validity. Retrospective study designs predominate, and control groups are often absent. Selection of prognostic timeframes frequently reflects data availability rather than clinical relevance. Established non-imaging prognostic scores are not consistently incorporated into imaging-based models, limiting opportunities to demonstrate incremental value.
Variability in performance metrics and reporting methods complicates comparison across models developed in different populations. These challenges collectively restrict the integration of artificial intelligence-driven prognostic tools into clinical workflows.
Artificial intelligence applied to imaging introduces new opportunities for prognostication in chronic liver disease, combining structural, functional and compositional data into predictive models. Applications span multiple modalities and endpoints, including survival, decompensation and cancer development. Despite promising performance across diverse use cases, limitations in dataset size, model transparency and study design constrain clinical adoption. Greater standardisation, larger datasets and integration with established prognostic frameworks remain central to advancing these approaches towards routine use.
Source: British Journal of Radiology
Image Credit: iStock
References:
Chouhan MD, McLean K, Thomas JA & Dowling J (2026) Artificial Intelligence prognostication of liver disease using imaging. British Journal of Radiology: tqag070.
