Liver and spleen volumes have important clinical implications across the spectrum of chronic liver disease. The liver volume reflects functional hepatic mass, while the spleen responds sensitively to portal pressure and enlarges predictably with advancing portal hypertension. Despite this relevance, volumetric assessments of these organs have historically been confined to research settings, as linear measurements are imprecise for geometrically complex organs and manual CT/MRI volumetry is too time-consuming for routine use [1-5].

Deep learning-based segmentation algorithms now perform fully automated liver and spleen volumetry from routine abdominal CT or MRI within seconds, at an accuracy comparable to expert manual segmentation. Gadoxetic acid-enhanced MRI extends this approach by combining volumetric and functional hepatic data in a single acquisition, simultaneously yielding the liver volume, spleen volume, and liverto-spleen signal intensity ratio as surrogates of hepatocellular function and portal pressure [7, 8]. Taken together, these advances position the liver and spleen volumes as scalable, workflow-integrated quantitative biomarkers applicable to large clinical populations.

A substantial body of evidence now documents the clinical utility of AI-derived volumetry across hepatic fibrosis staging, clinically significant portal hypertension (CSPH) detection, the prediction of liver-related outcomes as well as prognosis after hepatocellular carcinoma (HCC) treatment [7, 9-17]. This present narrative review synthesizes this body of evidence, discusses current technical and clinical limitations, and identifies priority areas for future validation and clinical implementation. The overall framework — from imaging input through AI segmentation to clinical application — is summarized in Fig. 1.

Early automated liver segmentation procedures relied on atlas-based registration and classical image processing methods, which were limited by sensitivity to morphological variation and pathological change. The introduction of convolutional neural networks, particularly encoder-decoder architectures such as U-Net and its variants, brought a step change in segmentation performance. These models learn hierarchical spatial features from large annotated training datasets and generalize well across varying liver conditions, including fatty liver, non-cirrhotic chronic liver disease, cirrhosis, and post-hepatectomy states [1, 18, 19]. For multi-organ applications, transformer-based pipelines such as TotalSegmentator now enable simultaneous segmentation of over 104 anatomic structures from a single CT scan, providing an infrastructure for large-scale opportunistic screening [6].

In a prior study, a validated deep learning algorithm demonstrated high segmentation accuracy on portal venous phase CT, with Dice similarity coefficient (DSC) values of 0.973 for the liver and 0.974 for the spleen. Volumetric measurements showed Bland- Altman 95% limits of agreement within ±5% across a range of liver conditions, and similar performance was observed on CT scans from external institutions [1]. Segmentation errors requiring manual correction occured in approximately 10-17% of cases, with a mean correction time of 17-37 seconds — well within routine clinical workflow tolerances. For nonenhanced CT and low-dose chest CT, nnU-Net- based models achieved comparable accuracy (DSC >0.96), broadening their applicability to opportunistic screening contexts where contrast-enhanced imaging may not be available [20, 21].

Gadoxetic acid-enhanced MRI may present additional segmentation challenges due to hepatobiliary phase signal intensity variation driven by flip angle, repetition time, and the degree of hepatocellular function. A dedicated deep learning algorithm for hepatobiliary phase MRI achieved a DSC of 0.953-0.981 for liver and spleen segmentation and simultaneously extracts the liver-to-spleen signal intensity ratio (LS-SIR). An adjusted LS-SIR (aLSSR) that corrects for MRI acquisition parameters enables comparable functional measurements across different scanner protocols, which is a prerequisite for multicenter clinical application [7, 8].

Establishing normative reference values is a prerequisite for translating volumetric measurements into clinical decisions. Two large-scale studies have provided complementary evidence for populationbased reference intervals using AI-based CT volumetry.

In a prior study using AI-based CT volumetry in 2,989 healthy adult liver donors in South Korea, population-based reference intervals were defined as follows: liver volume, 824.5-1700.0 cm³; spleen volume, 81.1-322.0 cm³; and liver-to-spleen volume ratio (LSVR), 3.96-13.78 [2]. These intervals were validated in an independent healthy cohort and compared against standard liver volume formulas derived from Asian and Western populations, with the AI-derived values showing superior agreement with measured volumes. The same study further demonstrated that quantile regression modeling of the 2.5th and 97.5th percentiles as a function of these anthropometric variables generates personalized reference intervals that outperform population-based thresholds in detecting abnormal volumes in patients with viral hepatitis [2]. A web-based calculator for implementing these personalized formulas is publicly available (https://aim-aicro.com/software/Calculators), enabling point-of-care application without specialized software.

In a separate study analyzing 8,901 screening CT examinations in the United States, AI-based spleen segmentation enabled derivation of weight-based splenomegaly thresholds that outperformed the traditional 13 cm linear criterion [22]. The mean spleen volume in this cohort was 216 cm³. Similarly, a weight-based volumetric threshold for hepatomegaly derived from this large asymptomatic screening cohort outperformed conventional craniocaudal linear measurements, which yield a sensitivity of only 71- 78% for volumetrically confirmed hepatomegaly [3].

Direct comparison of these cohorts highlights substantial geographic and ethnic variation in normal organ volumes — for example, mean spleen volumes differ meaningfully between the Korean donor cohort (194 cm³ in men, 149 cm³ in women) and the US screening cohort (216 cm³), such that a threshold derived from one population may systematically misclassify individuals from another [2, 22]. Because liver and spleen volumes also vary systematically with age, sex, height, and weight, population-based thresholds may misclassify individuals at the extremes of body habitus regardless of ethnicity.

Key findings for liver fibrosis evaluations across clinical applications are summarized in Table 1. Progressive hepatic fibrosis induces characteristic volumetric changes detectable on cross-sectional imaging. As fibrosis advances, the liver undergoes lobar redistribution — with relative hypertrophy of the caudate lobe and left lateral segments and atrophy of the right lobe — which is reflected in a rising liver segmental volume ratio (defined as the volume of Couinaud segments I-III relative to segments IV-VIII). Simultaneously, increasing portal hypertension drives progressive splenomegaly. The total liver volume alone is a poor discriminator of fibrosis stage, as it remains relatively preserved until advanced cirrhosis whereas the liver segmental volume ratio and spleen volume show strong, monotonic associations with both the histopathologic fibrosis stage and decompensation status across the full spectrum from no fibrosis to decompensated cirrhosis [17, 23, 24]. It is important to note, however, that the pattern and degree of lobar volumetric redistribution vary by etiology: alcohol-related and Hepatitic C virus (HCV)-related cirrhosis tend to produce more pronounced right lobe atrophy and caudate hypertrophy than Hepatitis B virus (HBV)-related disease, while cholestatic conditions such as primary sclerosing cholangitis may exhibit different segmental volume distributions. These etiology-specific morphological differences indicate that volumetric thresholds and ratios derived from predominantly HBV-related cohorts may not be directly transferable to other disease contexts without independent validation [25].

In a prior study of patients with chronic liver disease, an AI-measured spleen volume achieved area under the receiver operating characteristic curves (AUCs) of 0.82-0.87 for advanced fibrosis and cirrhosis detection, while the LSVR achieved AUCs of 0.82-0.88 [17]. A further report of a deep learning model measuring Couinaud segmental volumes and the spleen from contrast-enhanced CT performed comparably to radiologist manual measurements for predicting histopathologic cirrhosis (AUC 0.85 vs 0.86 for spleen volume; AUC 0.79 vs 0.83 for liver segmental volume ratio), with non-inferiority formally demonstrated [24]. It is important to note that these volumetric indices complement rather than replace existing non-invasive tools as they are independent of operator skill, can be extracted retrospectively from previously acquired CT data, and provide information not captured by serum-based scores such as Fibrosis-4 index (FIB-4) or aspartate aminotransferase-toplatelet ratio index (APRI). Whereas elastography directly measures liver stiffness as an upstream surrogate of fibrosis, volumetric indices reflect the downstream morphologic consequences of fibrosis — including lobar redistribution and splenomegaly — and are independent of operator skill and applicable retrospectively to previously acquired CT data. Volumetric assessment may be particularly useful when elastography is technically limited (e.g., obesity, ascites) or unavailable, positioning AI-based volumetry as a complementary tool in current fibrosis assessment pathways.

An important consideration in volumetric fibrosis assessments is the confounding effect of hepatic steatosis, which causes liver enlargement independent of fibrosis stage and may lead to overestimation of liver volume relative to fibrosis severity. AI-based measurements of liver volume combined with CT attenuation-based steatosis assessment — or MRIderived proton density fat fraction — allows this confounder to be quantified and corrected, thereby improving the specificity of volumetric fibrosis staging [26, 27]. Population-based automated liver fat quantification from non-enhanced CT and automated whole-liver MRI segmentation for proton density fat fraction and iron estimation have demonstrated that these measurements can be obtained at scale with an accuracy comparable to manual region-of-interest sampling [26, 27].

The spleen is exquisitely sensitive to portal hemodynamic changes and splenic congestion and extramedullary hematopoiesis driven by elevated portal venous pressure result in progressive and measurable splenomegaly. The spleen volume, therefore, serves as one of the most reliable imaging surrogates of portal hypertension severity, correlating more strongly with hepatic venous pressure gradient (HVPG) than the splenic diameter or platelet count alone [14, 28].

A validated CT-based diagnostic framework for CSPH integrates AI-measured spleen volume, assessed against personalized reference thresholds, with the presence of gastroesophageal varices (GEV), spontaneous portosystemic shunts (SPSS), and ascites on CT. In a previously studied cohort of 235 patients with chronic liver disease and HVPG measurement, personalized splenomegaly criteria outperformed fixed-threshold definitions, achieving an AUC of 0.91, in detecting CSPH [14]. A rule-in criterion combining personalized splenomegaly with at least one additional CT finding has achieved a specificity of 94.4-97.6% and a rule-out criterion based on the absence of all CT findings achieves a sensitivity of 94.5-98.2%. Applying these criteria, 72.8-78.7% of patients can be correctly classified without HVPG measurement, with misclassification rates of only 3-5.5%. This CT-based approach aligns conceptually with Baveno VII noninvasive testing recommendations [29] while offering the practical advantage of deriving all required data from a single routine CT examination. Whereas Baveno VII criteria require dedicated liver stiffness measurement and leave a substantial gray zone of indeterminate cases, CT-derived volumetric criteria can be obtained from imaging already acquired for other indications and may serve a complementary role in resolving such gray-zone cases, or function as an alternative non-invasive pathway when elastography is technically limited or unavailable.

External validation across diverse disease etiologies supports the generalizability of CT-based volumetric indices for CSPH diagnosis. In a prior cohort of 175 patients with HCC-related compensated cirrhosis and predominantly HCV- or alcohol-related disease, a volume index combining spleen volume and liver segmental volume ratio predicted CSPH with an AUC of 0.83, with comparable performance in an external validation cohort (AUC 0.87) [28]. Applying rule-in and rule-out thresholds could have averted 35.9% of HVPG measurements, and the association was independent of underlying etiology.

Beyond CSPH diagnosis, AI-measured spleen volume enables quantitative risk stratification for high-risk esophagogastric varices requiring treatment [11, 30]. An index combining spleen volume with platelet count in patients with HBV-related compensated cirrhosis demonstrates that a >1.63 cutoff value for the spleen volume-to-platelet ratio detects all high-risk varices with 100% sensitivity, while a cutoff of >3.78 achieves 69.4% sensitivity and 78.5% specificity for balanced performance [11]. Three-tier risk stratification based on this index predicts a cumulative 5-year variceal bleeding incidence of 0%, 1.0%, and 12.0% in low-, intermediate-, and high-risk groups respectively, enabling selective endoscopic surveillance in clinical practice. A representative case illustrating CT-based CSPH diagnosis is presented in Fig. 2A.

Liver and spleen volumes track closely with clinical disease severity across diverse etiologies and clinical settings. For example, as the Child-Pugh class advances from A to C, the liver volume decreases while the spleen volume increases, reflecting hepatocellular loss and worsening portal hypertension [25]. Consistent with this, liver and spleen volumetric indices have been shown to evolve in parallel with cirrhosis natural history in a study spanning compensated and decompensated disease — the liver segmental volume ratio captured the hemodynamic phase of compensated disease, and the standardized liver volume correlated inversely with Model for End-Stage Liver Disease (MELD) and HVPG, independently of fibrosis severity and etiology [31]. These observations provide the mechanistic basis for the prognostic associations described below.

At a single time point, AI-measured LSVR on CT independently predicts hepatic decompensation and transplantation-free survival in patients with compensated cirrhosis. In a prior report of 1,027 HBV-related compensated cirrhosis patients followed for a median of 116 months, an LSVR below 2.9 was associated with a 3-year decompensation incidence of 16.7% versus 2.5% above the threshold (P<0.001), independent of Child-Pugh and MELD [12]. Spleen volume alone also carries prognostic value: the CHESS1701 model incorporating spleen volume, cholinesterase, and hemoglobin (n=689, five centers) achieved a C-index exceeding 0.84 across training and two external validation cohorts, thereby outperforming Child-Pugh, MELD, Albumin-bilirubin grade (ALBI), and FIB-4, with a low-risk subgroup 3-year decompensation risk below 1%, comparable to Baveno VII criteria [9].

Beyond single time-point assessment, AI-based segmentation enables longitudinal prognostic monitoring. In 280 patients with advanced chronic liver disease (ACLD) followed for a median of 104 months, both baseline LSVR and LS-SIR automatically measured on the hepatobiliary phase of gadoxetic acid-enhanced MRI independently predicted decompensation in compensated ACLD (baseline LSVR: subdistribution hazard ratio 0.71) [8]. More importantly, serial measurement of ΔLSVR (the relative change in LSVR over one year) predicted both subsequent decompensation (sHR 0.54) and liverrelated death or transplantation (sHR 0.46). A ΔLSVR cutoff of ±13% successfully stratified patients into three clinically distinct risk groups, with a decrease exceeding 13% identifying a subgroup at near-zero decompensation risk over five years. These findings indicate that dynamic volumetric tracking captures disease trajectory information not available from static measurements.

AI-based liver and spleen volumetry informs outcome prediction across the spectrum of HCC management, from preoperative risk stratification for hepatic resection to prognostication in patients undergoing locoregional therapy. In the surgical setting, an adequate future liver remnant (FLR) volume is the cornerstone of safe hepatic resection, with minimum FLR thresholds of 20-25% for patients with normal hepatic reserve and 40% or more in the setting of cirrhosis or significant fibrosis — as defined by the International Study Group of Liver Surgery (ISGLS) [32]. Accurate CT volumetry of the total liver and planned remnant is therefore integral to preoperative planning for a major hepatectomy and AI-based automated segmentation now makes this measurement practical in routine clinical settings. Beyond remnant volume alone, however, accumulating evidence indicates that quantitative assessment of the spleen and of regional hepatic function provides additional and independent prognostic information.

On preoperative CT, the spleen volume serves as a surrogate of portal hypertension severity that compromises remnant liver regeneration. In 317 consecutive early-stage HCC patients undergoing resection, the spleen volume divided by body surface area (SV/BSA)was significantly higher in those who developed post-hepatectomy liver failure (PHLF) (148 vs 112 cm³/m², P<0.001; AUC 0.663), and an SV/BSA above 146 cm³/m² independently predicted overall survival (HR 3.935) [16]. Building on this, the CTbased CSPH — defined by personalized splenomegaly criteria combined with the presence of GEV, SPSS, or ascites — captures the full spectrum of adverse posthepatectomy outcomes in patients with ACLD and HCC. In a population of 593 patients undergoing hepatectomy, CT-based CSPH was independently associated with severe PHLF (odds ratio 7.672), hepatic decompensation (sHR 4.518), and liverrelated death or transplantation (sHR 2.756) [15]. Importantly, the CT-based CSPH outperformed conventional CSPH criteria in augmenting the predictive performance of established PHLF models including the EASL algorithm³3 and Wang's model34, and its more conservative classification rate (7.9% vs 12.5%) reflects superior specificity [15].

Whereas CT volumetry quantifies anatomic reserve, gadoxetic acid-enhanced MRI uniquely enables simultaneous assessment of remnant liver volume and hepatocellular function through the aLSSR of the planned remnant region. In 1,760 HCC patients across two institutions, MRI-based nomograms incorporating remnant liver volume, aLSSR, spleen volume normalized to body surface area, and standard clinical variables achieved AUCs of 0.81-0.84 for symptomatic PHLF prediction, thereby outperforming existing models based on CT volumetry or serum markers alone [13]. The remnant-specific model outperformed the whole-liver model, underscoring the importance of region-specific functional assessment. A representative case illustrating MRI-guided PHLF risk assessment is described in Fig. 2B.

Beyond perioperative risk, preoperative spleen volume also predicts long-term oncologic outcomes that reflect the underlying cirrhotic field rather than the resected tumor. Among 301 HCC patients followed for a median of 48.8 months after curative resection, spleen volume was the only volumetric index to differ significantly between those with and without late recurrence (median 321.4 vs 268.9 cm³), whereas liver-related indices did not [10], consistent with late recurrence representing de novo tumor development driven by the cirrhotic field. Supporting this interpretation, a larger CT-measured spleen volume independently predicts de novo HCC development in compensated chronic liver disease: in 584 HBVrelated patients followed for a median of 92 months, a threshold of 532 mL identified a high-risk subgroup with a 7-year cumulative HCC incidence of 17.2%, independent of FIB-4 [35]. The prognostic relevance of spleen volume extends beyond the surgical setting: in HCC patients undergoing transarterial chemoembolization (TACE), AI-measured spleen volume outperformed two-dimensional diameter for predicting overall survival and decompensation risk [36]. Together, these findings position spleen volume as an integrative surrogate of the portal hypertensive and hepatocarcinogenic milieu across the full spectrum of HCC management.

Opportunistic imaging refers to the systematic extraction of clinically relevant biomarkers from CT or MRI examinations acquired for unrelated primary indications, thus leveraging data already present in the PACS without additional imaging, cost, or patient burden. In this regard, the liver and spleen volumes are among the most compelling of the opportunistic biomarkers as they are quantitative, reproducible, and derivable from portal venous phase CT (the most commonly performed abdominal CT protocol), as well as from non-enhanced CT and low-dose chest CT using modern nnU-Net-based algorithms [37, 38]. The potential clinical scale and impact is substantial given that hundreds of millions of abdominal CT examinations are performed annually worldwide, the majority of which currently yield no volumetric organ data despite containing the raw information necessary for its extraction. Population-based data underscore this potential. In large general population cohorts, incidentally detected splenomegaly on routine CT or MRI was reported to be associated with substantially elevated risks of hematologic cancer and liver disease, with these risks rising sharply above spleen volume thresholds of 400-500 mL and demonstrating that volumetric findings carry actionable prognostic information beyond their primary clinical indication [39].

In practice, a human-in-the-loop workflow, in which automated segmentation results are reviewed and corrected by a radiologist or trained technician in under two minutes, provides a pragmatic balance between full automation and clinical reliability. PACS-integrated pipelines can trigger background processing automatically upon CT or MRI acquisition, delivering verified volumetric outputs at the time of radiological reporting. Web-based calculators incorporating personalized reference intervals, CSPH criteria, and PHLF nomograms translate raw volumetric measurements into actionable clinical risk estimates at the point of care [2, 13, 14]. Multi-organ AI pipelines such as TotalSegmentator further enable the simultaneous extraction of liver volume, spleen volume, body composition parameters, and other imaging biomarkers from a single examination, positioning liver and spleen volumetry within a broader framework of comprehensive CT-based phenotyping [6, 37].

Despite impressive segmentation accuracy under standard conditions, current deep learning algorithms retain meaningful failure modes: CT performance degrades with severe artifacts, extensive hepatic cysts, post-surgical distortion, or non-portal venous phase acquisitions, while gadoxetic acid-enhanced MRI requires aLSSR-based parameter correction with residual inter-scanner variability limiting multiinstitutional deployment [1, 7]. Recent nnU-Net-based models for non-enhanced CT represent meaningful progress, though comprehensive validation across protocols and scanner types remains ongoing [21].

Furthermore, the majority of published evidence derives from single-center retrospective cohorts. Whether thresholds derived from predominantly HBVrelated cohorts generalize to metabolic dysfunctionassociated steatotic liver disease (MASLD), alcoholrelated, or autoimmune etiologies remains uncertain, particularly as hepatic steatosis independently enlarges liver volume, potentially confounding fibrosis-stage classification [25]. In addition, the relative clinical positioning of volumetric indices versus elastographybased tests has not been directly evaluated in the same cohort against HVPG or liver biopsy. Finally, while volumetric indices have demonstrated strong associations with clinical outcomes, no randomized controlled trial has yet demonstrated that volumetryguided decision-making improves patient outcomes compared to standard care, and this is a critical evidence gap for guideline integration. Large-scale multinational prospective validation, encompassing the full spectrum of chronic liver disease etiologies and diverse imaging protocols, is a prerequisite for guideline-level adoption of AI-derived volumetric biomarkers in global clinical practice.

Several directions will shape the next phase of AI- based liver and spleen segmentation and its clinical application. Segmentation refinement, including Couinaud segment-level volumetry, intrahepatic vascular quantification, and parenchymal texture analysis, will extract richer biological information from existing imaging data. Integration with serum fibrosis markers, elastography, or clinical scores into multimodal models has already shown additive discriminative value [40], with further fusion with genomic or electronic health record data enabling more granular phenotyping. Longitudinal volumetric monitoring may enable dynamic risk reassessment and treatment response evaluation, [8] while foundation models and multi-organ pipelines move toward phaseagnostic, modality-agnostic volumetry for large-scale opportunistic screening [19, 41]. Finally, integration into structured radiology reporting, supported by regulatory-cleared AI tools, will be essential for translating evidence into routine clinical practice.

AI-based automated segmentation has transformed liver and spleen volumetry from a research-only measurement system into a scalable, clinically actionable biomarker applicable across the full continuum of chronic liver disease management. From noninvasive fibrosis staging and CSPH detection to the prediction of hepatic decompensation, HCC recurrence, and post-hepatectomy liver failure, AIderived volumetric indices consistently provide independent prognostic information beyond conventional clinical and laboratory markers. The integration of gadoxetic acid-enhanced MRI extends this capability by combining structural and functional hepatic assessment in a single acquisition. As opportunistic imaging pipelines mature and multiorgan AI platforms become embedded in routine radiology workflows, liver and spleen volumes are poised to become standard components of the quantitative imaging report and thereby inform clinical decisions at the point of care. In current practice, these indices are best positioned as complementary tools that augment established non-invasive tests such as elastography, serum-based fibrosis scores, and clinical prognostic systems, with the potential to serve as an alternative non-invasive pathway when those tests are technically limited or unavailable. Prospective multicenter validation across diverse disease etiologies, alongside regulatory approval and structured workflow integration, will be the critical next steps toward realizing this potential in global clinical practice.