Artificial intelligence in expanding hepatic resection boundaries: Integrating portal flow modulation and regenerative strategies

Table 1 Artificial intelligence and machine learning applications in liver surgery

Application area	AI/ML model type	Data source	Clinical impact	Performance metric
Liver segmentation/volumetry	CNN (U-Net, variants)	CT, MRI	Automated FLR measurement, resection planning	Dice coefficient > 0.95
FLR function prediction	Radiomics, ML classifiers	MRI (Gd-EOB-DTPA), CT	Predicts PHLF and functional margins	AUC 0.82-0.94
Outcome prediction (PHLF, complications)	Gradient boosting, Light GBM	EHR, imaging	Individualized risk, clinical DSS	AUC 0.82-0.94
Tumor segmentation/classification	Deep CNN	CT, MRI	Automated detection, margin planning	Accuracy > 93%
Intraoperative decision support	Explainable ML, AR	Video, segmentation	Real-time guidance, workflow efficiency	Not routinely quantified

CNN: Convolutional neural network; FLR: Future liver remnant; EOB-DTPA: Gadoxetic acid (gadoxetate disodium; ethoxybenzyl-diethylenetriamine pentaacetic acid); PHLF: Post-hepatectomy liver failure; DSS: Decision support system; AUC: Area under the curve; CT: Computed tomography; AI: Artificial intelligence; ML: Machine learning; MRI: Magnetic resonance imaging; AR: Augmented reality.

Table 2 Comparison of portal flow modulation techniques

Technique	Mechanism	Median FLR hypertrophy	Hypertrophy time	Main risks	Indications
Portal vein embolization (PVE)	Portal inflow occlusion	37%-40%	3-6 weeks	Low (abscess, thrombosis)	Insufficient FLR, most common bridge
Liver venous deprivation	PVE + hepatic vein embolization	50%-70%	2-4 weeks	Slightly higher dysfunction risk	Inadequate hypertrophy after PVE
Associating liver partition and portal vein ligation	Portal ligation + parenchymal transection	60%-80%	7-10 days	Higher: Infection, liver failure	Unresectable with rapid hypertrophy need

FLR: Future liver remnant.

Table 3 Molecular pathways and biomarkers in liver regeneration

Pathway/factor	Main functional role	Clinical or experimental marker	Relevance in regeneration
TNF-α/IL-6	Initiation phase, hepatocyte priming	Serum TNF-α, IL-6	Predicts early regenerative activity
HGF/Met	Proliferation, cell signaling	Serum HGF	Predicts regenerative capacity
Wnt/β-catenin	Proliferation, stem cell activation	Nuclear β-catenin	Regulates proliferation vs differentiation
mTORC1	Controls hypertrophy vs hyperplasia	Phospho-S6K, mTOR gene	Promotes protein synthesis for regeneration
MicroRNAs (e.g., miR-21, miR-122)	Transcriptional modulation	Circulating miRNA signatures	Biomarker for regeneration or injury

TNF-α: Tumor necrosis factor-α; IL-6: Interleukin-6; HGF: Hepatocyte growth factor.

Table 4 Radiomics features predicting surgical outcomes

Outcome	Imaging modality	Key radiomics features	Best classifier	Performance (AUC/accuracy)
Recurrence (CRLM)	EOB-MRI	VIBE_FA10 textural metrics	KNN	AUC 0.91, Acc. 93%
Tumor front growth (exp. vs inf.)	CT	7 textural features (GLCM, etc.)	KNN	Acc. 97%, Sens. 90%, Spec. 100%
Tumor budding	Contrast MRI	11 textural features (arterial phase)	KNN	Acc. 95%, Sens. 84%, Spec. 99%
Early recurrence	CT	Tumor quality and quantity model	ML ensemble	AUC 0.83
Macrovesicular steatosis (donor)	CT	7 selected radiomic features	Logistic regression	AUC 0.87

CRLM: Colorectal liver metastases; EOB-MRI: Gadoxetic acid-enhanced magnetic resonance imaging; GLCM: Gray-level co-occurrence matrix; KNN: K-nearest neighbors; AUC: Area under the curve; CT: Computed tomography; ML: Machine learning; Acc: Accuracy; Sens: Sensitivity; Spec: Specificity; MRI: Magnetic resonance imaging.

Table 5 Ethical, regulatory, and implementation challenges

Challenge area	Description	Example/impact
Algorithmic bias	Incomplete data, non-generalizable models	Less accurate for underrepresented groups
Transparency/explainability	Black-box predictions frustrate trust	Difficult to validate medical rationale
Generalizability	Overfit to single-center/population data	Poor results in new settings
Data privacy/regulation	Compliance with HIPAA/GDPR	Re-identification risks in cloud-based AI
Clinical validation	Lack of RCTs/prospective evidence	Hinders regulatory approval, uptake

HIPAA: Health insurance portability and accountability act; GDPR: General data protection regulation; AI: Artificial intelligence; RCT: Randomized controlled trial.

Table 6 Summary of study quality across major evidence categories

Category	Typical study design	Sample size range	Key findings	Level of evidence	Risk of bias (summary)
AI segmentation/volumetry	Technical validation, retrospective imaging datasets	40-1200 scans	Dice 092-0.97 for liver segmentation	IV	High risk-internal validation only, curated datasets, protocol heterogeneity
ML prediction models for PHLF	Retrospective multicenter or single-center	300-25000 patients	AUC 0.82-0.94 for PHLF prediction	III-IV	Moderate risk -class imbalance, unblinded outcome measurement, overfitting risk
Comparative PVE/LVD/ALPPS studies	Retrospective cohorts, meta-analyses	60-1800	LVD > PVE hypertrophy; ALPPS fastest hypertrophy	II-III	Moderate risk -selection bias, inconsistent endpoints, non-standard hypertrophy intervals
Radiomics prognostic studies	Retrospective, mostly single-center	40-300	Predict recurrence, FLR dysfunction	IV	High-to-moderate risk -small samples, overfitting, rare external validation
Preclinical regenerative biology	Rodent and in vitro	n = 6-60 animals	Pathway-level mechanistic insights	V	Low-to-moderate risk-mechanistic but non-clinical

FLR: Future liver remnant; PVE: Portal vein embolization; LVD: Liver venous deprivation; ALPPS: Associating liver partition and portal vein ligation for staged hepatectomy; ML: Machine learning; AI: Artificial intelligence; AUC: Area under the curve; PHLF: Post-hepatectomy liver failure.

Table 7 Artificial intelligence-supported decision algorithm

Predicted parameter	Threshold	Suggested clinical action
Predicted PHLF risk	> 20%-25%	Avoid major hepatectomy; consider LVD or two-stage strategies
	10%-20%	Prefer LVD over PVE; optimize FLR function, nutrition
	< 10%	PVE adequate; proceed with resection after hypertrophy
Predicted hypertrophy insufficient after PVE (model forecast)	< 25% expected growth	Prefer primary LVD
Tumor progression risk during waiting period	High	Consider ALPPS or accelerated LVD
Functional FLR prediction (radiomics/MRI)	Below functional cutoffs	Avoid ALPPS; prefer staged or non-surgical options

LVD: Liver venous deprivation; ALPPS: Associating liver partition and portal vein ligation for staged hepatectomy; MRI: Magnetic resonance imaging; FLR: Future liver remnant; PVE: Portal vein embolization.

Table 8 Comparative characteristics of portal vein embolization, liver venous deprivation, and associating liver partition and portal vein ligation for staged hepatectomy

Feature	PVE	LVD	ALPPS
Typical study design	Retrospective cohorts	Retrospective + prospective pilot	Retrospective ALPPS registry
Sample size range	0-50	0-40	0-70
Median hypertrophy (%)	0-37	0-50	0-60
Time to hypertrophy (days)	2-21	8-10	0-7
Failure-to-resect rate (%)	0-5	0	0
90-day morbidity (%)	0-10	5-10	5-35
90-day mortality (%)	0	0	2-8
Key limitations	Collateral formation	More complex, higher dysfunction risk	Highest morbidity and selection bias

PVE: Portal vein embolization; LVD: Liver venous deprivation; ALPPS: Associating liver partition and portal vein ligation for staged hepatectomy.