AI applications in liver interventions: from routine shortcomings to AI solutions.
| Phase | Clinical application | Conventional method | AI methodology | AI advantage | Reference(s) |
|---|---|---|---|---|---|
| Pre-procedural | Predicting TACE response | Visual CT/MRI & BCLC staging. High inter-observer variability; fails to capture sub-visual tumor heterogeneity. | Multimodal models using radiomics, DL, and clinical data (ALBI, BCLC, AFP). | AUROC > 0.85. Improves patient selection and avoids futile procedures. | [6, 28, 37, 38] |
| Non-invasive PHT assessment | Invasive HVPG measurement. Procedural risk and requirement for highly specialized expertise. | Radiomics and DL models (e.g., aHVPG) are analyzing CT features of the liver and spleen. | Non-invasively estimate HVPG to stratify risk and guide TIPS candidacy. | [1, 4, 53] | |
| Predicting post-TIPS complications | Clinical scores (MELD, Child-Pugh). Limited predictive power for post-TIPS OHE. | Radiomics, ANNs, and various ML models. | Accurately forecast the risk of OHE for counseling. | [3, 4] | |
| Predicting PVE success | 2D/3D CT volumetry. Volume does not always equal function; difficult to predict actual hypertrophy kinetics. | Multimodal models using Statistical Shape Models to quantify 3D liver anatomy. | Forecast FLR hypertrophy to optimize surgical planning. | [7] | |
| Intra-procedural | Treatment simulation | Standard anatomical landmarks. Fails to account for heat-sink effects or perfusion-based boundaries. | DL models to predict ablation zones and simulate Y-90 radioembolization dosimetry. | Optimize probe placement and increase quantitative accuracy of dosimetry. | [24, 39–41] |
| Image quality improvement | Conventional imaging filters. High radiation dose or poor visualization due to artifacts. | DLR for dose reduction; DL to reduce metal artifacts or generate synthetic DSA. | Lower radiation dose; improve visualization and safety during procedures. | [19, 24, 47] | |
| Post-procedural | Longitudinal monitoring of HCC | mRECIST criteria. Does not account for dynamic metabolic changes or internal necrosis patterns. | DL (Transformers) using multi-time-point MRI data to track tumor changes. | More accurate prognostic stratification than diameter-based criteria. | [28] |
| Detecting tumor recurrence | Manual surveillance review. Potential for human error in identifying subtle early progression. | ML, radiomics, and CNNs. | Automated and early detection of LTP after ablation. | [24, 46] |
AI: artificial intelligence; TACE: transarterial chemoembolization; BCLC: Barcelona Clinic Liver Cancer; DL: deep learning; ALBI: albumin-bilirubin; AFP: alpha-fetoprotein; AUROC: area under the receiver operating characteristic curve; PHT: portal hypertension; HVPG: hepatic venous pressure gradient; TIPS: transjugular intrahepatic portosystemic shunt; MELD: model for end-stage liver disease; OHE: overt hepatic encephalopathy; ANNs: artificial neural networks; ML: machine learning; FLR: future liver remnant; Y-90: Yttrium-90; DLR: DL reconstruction; DSA: digital subtraction angiography; HCC: hepatocellular carcinoma; mRECIST: Modified Response Evaluation Criteria in Solid Tumors; CNNs: convolutional neural networks; LTP: local tumor progression.