Differentiating multilocular hepatic cysts from mucinous cystic neoplasms: characteristic imaging signs and a machine learning diagnostic framework

Introduction

Hepatic cystic lesions represent a heterogeneous group of conditions, among which multilocular hepatic cysts (MHCs) and mucinous cystic neoplasms (MCN) exhibit distinct pathological features and clinical trajectories. MHCs are benign lesions, most commonly identified incidentally in asymptomatic women, with a reported prevalence of 2.5% to 5% (1,2). In MHCs, intracystic hemorrhage, although a rare complication, may result in severe abdominal pain (3). These lesions typically appear as multilocular cystic structures characterized by thin walls and uncomplicated internal architecture; however, the diversity in imaging features can complicate diagnostic interpretation (1). In contrast, MCNs account for less than 5% of all cystic liver lesions, also predominantly affecting women, and are associated with a potential risk of malignant transformation (4,5). On imaging, MCNs generally present as solitary cystic lesions with thickened walls, mural nodules, and septal thickening (6,7).

Differentiating MHCs from MCNs based on imaging characteristics remains challenging due to substantial radiological overlap (8). The diagnosis of MHC relies primarily on imaging modalities, with laboratory testing contributing to the assessment of parasitic infections and malignant potential (9). In contrast, the presence of ovarian-type stroma is required for the histopathological confirmation of MCN (10). Several studies have identified imaging characteristics suggestive of MCNs, including septal thickness >3 mm, the presence of solid nodules, adjacent bile duct dilatation, and fewer than 3 cystic lesions (11,12). However, imaging alone is frequently insufficient for definitive diagnosis (13). MHCs may develop imaging features suggestive of malignancy during follow-up, further complicating the diagnostic process (14). Additionally, 20% to 50% of MCNs are inaccurately diagnosed prior to surgical intervention, potentially resulting in inappropriate treatment strategies or increased surgical risk (12).

Machine learning techniques have demonstrated advantages over traditional statistical methods in feature selection and diagnostic model development. Xiao et al. (2024) used a random forest (RF) algorithm to identify four key magnetic resonance imaging (MRI) features, achieving an area under the curve (AUC) of 0.982 (15). Hardie et al. (2022) developed a machine learning-based classification system that achieved 93.5% accuracy in distinguishing MCNs from benign hepatic cysts (16). The Boruta algorithm facilitates the identification of relevant features by excluding non-contributory variables, thereby minimizing multicollinearity (17). Least absolute shrinkage and selection operator (LASSO) regression incorporates penalization to reduce overfitting, enabling more interpretable and predictive models (18). RF, decision tree, and extreme gradient boosting (XGBoost) models can manage complex nonlinear associations and high-dimensional data structures, resulting in enhanced diagnostic accuracy and model robustness. Therefore, this study integrates these methods to construct a reliable diagnostic tool for differentiating MCNs from MHCs.

This retrospective study aimed to identify imaging features with diagnostic value for distinguishing MHCs from MCNs, construct and assess predictive models using various machine learning algorithms, and improve the accuracy of differential diagnosis, to support clinicians in managing hepatic cystic lesions more effectively. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2595/rc).

Methods

Study participants

Data were collected from participants at Tianjin Medical University General Hospital and Tianjin Fifth Central Hospital between July 2014 and April 2023. Inclusion criteria comprised the availability of preoperative computed tomography (CT) and/or MRI with identifiable multilocular hepatic imaging features. Exclusion criteria included the presence of cystic lesions communicating with the bile ducts, confirmed diagnosis of polycystic liver disease, and imaging studies that did not meet diagnostic quality standards. A total of 60 cases of MHC and 26 cases of MCN, all confirmed by surgical pathology, were enrolled. All patients included in the study underwent preoperative CT and/or MRI (Figure 1). This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study was approved by the Clinical Trial Ethics Committee of Tianjin Fifth Central Hospital (No. WZX-EC-KY2024031). This study is a retrospective analysis of medical records, and data anonymization is only used for undergraduate research. Therefore, the ethics committee waived the requirement for written informed consent.

Figure 1 Flowchart illustrating the study design. CT, computed tomography; LASSO, least absolute shrinkage and selection operator; MCN, mucinous cystic neoplasm; MHC, multilocular hepatic cyst; MRI, magnetic resonance imaging.

Imaging equipment and protocols

For CT imaging, participants were positioned in the supine position, and the scanning range was set to include the entire liver. Following image acquisition, thin-slice reconstruction and multiplanar reformation were conducted. For contrast-enhanced imaging, iodixanol was administered via the antecubital vein, and images were obtained during arterial, portal venous, and equilibrium phases.

MRI examinations were performed with participants in the supine position, covering the region from the diaphragmatic dome to the lower poles of both kidneys. The MRI protocol included axial T2-weighted imaging, axial T1-weighted imaging using a fast gradient-echo in-phase and opposed-phase water-fat sequence, coronal T2-weighted imaging using a single-shot fast spin-echo sequence, and axial diffusion-weighted imaging. Contrast-enhanced MRI was performed using a three-dimensional volumetric interpolated breath-hold T1-weighted gradient-echo sequence to acquire arterial, portal venous, equilibrium (transitional), and hepatobiliary phase images. Imaging parameters were standardized across all scanners to ensure consistency and reproducibility (Appendix 1).

Imaging feature analysis

Imaging analysis and quantitative measurements were performed independently by two radiologists who were blinded to clinical and pathological data. In instances of disagreement, consensus was reached through discussion. For participants with multifocal lesions, the largest lesion was selected for primary analysis. When smaller cysts were observed within a larger cystic structure, features of the dominant cyst were assessed. The following imaging features were evaluated: (I) overall lesion characteristics included location (left/right lobe), lesion type (solitary/multiple), number of lesions (≥20 or not), and overall lesion shape (round/lobulated). (II) Cyst wall characteristics included signal uniformity of cyst wall (homogeneous vs. heterogeneous), cyst wall thickness (≥2 vs. <2 mm), presence of mural nodules or solid components (defined as focal protrusion or segmental/diffuse wall thickening measuring ≥10 mm; yes/no), and presence of calcification (yes/no). (III) Septal characteristics included number of septa (≥5 vs. <5), septal thickness (≥3 vs. <3 mm), septal location [peripheral vs. central (any central involvement was classified as “central” regardless of additional peripheral septa)], and septal appearance (smooth vs. irregular). (IV) Indirect imaging signs included vascular displacement sign (defined as deep indentation of the cystic lesion due to vascular compression; present/absent), ribbon sign (defined as a septum with one end attached to the cyst wall and the other end free-floating; present/absent), biliary dilation sign (present/absent), septum-intersection triangular sign (defined as a “triangular” or “wedge-shaped” soft tissue density at the junction of septa or between the septum and cyst wall; present/absent), and vascular compression sign (present/absent).

Statistical methods

Comparative analyses between patients diagnosed with MHCs and those with MCNs were assessed using the t-test or Mann-Whitney U test for continuous variables based on the results of the Anderson-Darling normality test. Categorical variables were compared using the chi-squared test or Fisher’s exact test, as appropriate. Spearman’s correlation analysis was conducted to assess inter-variable relationships, and the results were visualized using a heatmap. Variables with a correlation coefficient (|r|) ≥0.8 were considered collinear and excluded from subsequent analyses to maintain model stability.

Three feature selection methods were applied to non-collinear variables: Boruta algorithm, LASSO regression, and univariate/multivariate logistic regression analyses. Diagnostic performance was assessed for variables selected by at least two of these methods.

Four clinical diagnostic models were constructed based on two sets of features: (I) variables identified by all three selection methods and (II) variables selected by at least two methods. The models were built using logistic regression, RF, decision tree, and XGBoost. The diagnostic performance of each model was assessed using the AUC. The highest-performing model was interpreted using Shapley Additive exPlanations (SHAP). Calibration curves and decision curve analysis (DCA) were generated to assess model calibration and clinical utility.

The association between the ribbon sign and intracystic hemorrhage was further examined using Firth logistic regression, adjusting for age and sex. All statistical analyses were conducted using R software (version 4.3.3; R Foundation for Statistical Computing, Vienna, Austria, 2024). A two-tailed P<0.05 was considered statistically significant.

Results

A total of 86 patients were included in this study, consisting of 60 patients with MHC and 26 with MCN. Compared to patients with MCN, those with MHC demonstrated a significantly higher median age (65.00 vs. 56.50 years, P<0.001), a higher incidence of intracystic hemorrhage (P=0.044), and more frequently presented with ≥20 hepatic cystic lesions (P=0.026). Lesions in the MHC group were more frequently located in the right hepatic lobe (P=0.018) and occurred more commonly as multiple lesions (P<0.001).

In contrast, imaging features more frequently observed in the MCN group included ≥5 septa (P=0.001), cyst wall thickness ≥2 mm (P=0.004), solid component thickness ≥10 mm (P=0.003), centrally located septa (P<0.001), irregular septa (P<0.001), and the presence of the septum-intersection triangular sign (P=0.013). However, no statistically significant difference was observed in septum cyst wall relationship (P=0.105, Table 1).

Spearman’s correlation analysis identified a correlation coefficient greater than 0.8 between the ribbon sign and intracystic hemorrhage, indicating potential collinearity (Figure S1). Variables exhibiting multicollinearity were excluded from further analyses to maintain model stability.

Three methods were used for the selection of morphological variables: Boruta, LASSO, and univariate and multivariate logistic regression. Variables selected by at least two of the methods included lesion type, septum appearance, ribbon sign, cyst wall thickness, septum-intersection triangular sign, age, and septum position (Figure 2, Table 2). Among these, lesion type, septum appearance, and ribbon sign were identified by all three methods, suggesting their relative importance among the morphological features. Representative imaging features of MHC and MCN are presented in Figures 3,4, while the detailed annotations of selected features are illustrated in Figure 5.

Figure 2 Feature selection results using Boruta and LASSO methods. Variables selected in the Boruta algorithm included lesion type, ribbon sign, age, septum appearance, septum position, septum number, presence of mural nodules or solid component thickness ≥10 mm, lesion location, septum-intersection triangular sign, cyst wall thickness, and lesion number (A). Green boxplots represent confirmed attributes, yellow represent tentative attributes, red represent rejected attributes, and blue represent shadow attributes. LASSO results are presented with the cross-validation curve (B), regression coefficient path plot (C), and a summary of the selected variables and their model impact (D). The optimal precision model identified 11 variables and was selected for use in this study, while the most parsimonious model identified four variables. LASSO-selected variables included lesion type, septum appearance, septum thickness, cyst signal uniformity, ribbon sign, age, cyst wall thickness, septum-intersection triangular sign, septum position, vessel compression, and sex. LASSO, least absolute shrinkage and selection operator; MCN, mucinous cystic neoplasm; MHC, multilocular hepatic cyst.

Figure 3 MRI and CT findings in a patient with MHC. Imaging from a 60-year-old male diagnosed with histopathologically confirmed MHC. (A-D) MRI findings: the lesion demonstrated heterogeneous low signal intensity on T1-weighted images and relatively high signal on T2-weighted images (A,B). The ribbon sign and septum were visible, with septa arising from the cyst wall (C,D). The T2-FS sequence showed high signal intensity in the cystic component and low signal intensity in the septa and ribbon sign (A,D). (E-H) CT findings: unenhanced CT demonstrated a solitary cyst in the left hepatic lobe with homogenous hypoattenuation (E) and no contrast enhancement across phases (F-H). CT, computed tomography; MHC, multilocular hepatic cyst; MRI, magnetic resonance imaging; T2-FS, T2-weighted fat-suppressed

Figure 4 MRI and CT findings in a patient with MCN. Imaging from a 75-year-old female with histopathologically confirmed MCN. (A) Unenhanced CT indicated an oval, uniformly hypodense cyst in the right hepatic lobe. (B,C) MRI demonstrated homogeneously high signal on. T2-weighted imaging low signal septa. (D) T1-FS post-contrast images revealed enhancement of septa and pericystic tissue, with no enhancement of the cyst contents. (E-H) Post-contrast T1-weighted imaging demonstrated persistent peripheral enhancement across all phases and heterogeneous enhancement consistent with intracystic hemorrhage. CT, computed tomography; MCN, mucinous cystic neoplasm; MRI, magnetic resonance imaging; T1-FS, T1-weighted fat-suppressed.

Figure 5 Indirect radiological signs evaluated in the study. (A) Ribbon sign on the T1-FS MRI: low signal septum arising from the cyst wall (red arrow). (B) Vascular displacement sign (red arrow) and vessel compression sign (blue square arrow) on CT: characterized by vascular indentation or deformation due to cyst compression. (C) Biliary dilation indicated by fluid-like tubular structures near the cyst on unenhanced CT (red arrow). (D) Septum-intersection triangular sign on T1-FS MRI: high signal triangular region at septal junctions (red arrow). (E) Septa originating from indented regions of the cyst wall. (F) Septa originating from flat regions of the cyst wall. CT, computed tomography; MRI, magnetic resonance imaging; T1-FS, T1-weighted fat-suppressed.

Lesion type demonstrated favorable diagnostic performance, with a sensitivity of 0.731 and specificity of 0.900. Septum position indicated balanced diagnostic characteristics, with a sensitivity of 0.538 and specificity of 0.867. In contrast, septum appearance (sensitivity: 0.308, specificity: 0.983), cyst wall thickness (sensitivity: 0.269, specificity: 0.967), and the septum-intersection triangular sign (sensitivity: 0.231, specificity: 0.967) were associated with high specificity but limited sensitivity. The ribbon sign exhibited high sensitivity (0.962) but relatively low specificity (0.267) (Table 3). Among all variables, lesion type achieved the highest diagnostic accuracy, with an AUC of 0.82 (Figure 6A).

Figure 6 Diagnostic performance of imaging features and machine learning models, with SHAP interpretation. ROC curves are presented for independent features and diagnostic models constructed using logistic regression, RF, decision tree, and XGBoost. SHAP analysis illustrates the contributions of each feature to the XGBoost model’s output. AUC, area under the curve; RF, random forest; ROC, receiver operating characteristic; SHAP, SHapley additive exPlanations; XGBoost, extreme gradient boosting.

Logistic regression, RF, decision tree, and XGBoost models were created to assess the diagnostic performance of the selected imaging features. The XGBoost model demonstrated the highest diagnostic performance, with an AUC of 0.849, when limited to the three indicators identified by all three feature selection methods (lesion type, septum appearance, and ribbon sign) (Figure 6B). To satisfy the empirical requirement of 10 events per variable, the logistic regression model was restricted to these three indicators. SHAP values were calculated and ranked for XGBoost models. In the three-indicator model, with SHAP values of 0.261 (Figure 6B).

Upon expansion to include six variables selected by at least two methods, adding cyst wall thickness, septum-intersection triangular sign, and septum position, the diagnostic reliability of the XGBoost model further improved, achieving an AUC of 0.905 (Figure 6C). Calibration curves and DCA were generated for these models (Figures S2 and S3). In the six-indicator models, lesion type still contributed the most to the predictive output, with SHAP values of 0.251 (Figure 6C).

In the SHAP importance swarm plot for the six-indicator model, lesions characterized as multilocular, as well as those with a ribbon sign (SHAP value =0.071), regular septum appearance (SHAP value =0.042), peripheral septal location (SHAP value =0.066), absence of the septum intersection triangular sign (SHAP value =0.013), and thin cyst walls (SHAP value =0.001), were more likely to be classified as MHC (Figure 6C). Conversely, the absence of these features was associated with an increased likelihood of MCN classification.

Firth logistic regression, adjusted for age and sex, demonstrated a significant association between the ribbon sign and the presence of intracystic hemorrhage [odds ratio (OR) =608.46, P<0.001]. The interaction between intracystic hemorrhage and disease type (MHC or MCN) was not significantly associated with the ribbon sign [OR =0.03, 95% confidence interval (CI): 0.00–7.38, P=0.272, Table 4].

Discussion

Study overview

This study aimed to improve the diagnostic differentiation between MHCs and MCNs by identifying key radiological features and applying multiple machine learning algorithms to develop predictive models. The imaging feature set evaluated in this study was constructed from three complementary sources to ensure comprehensiveness and relevance: (I) well-established features from prior literature (e.g., lesion number, mural nodule, septal thickness); (II) features derived from radiological principles (e.g., septal location and appearance); and (III) exploratory features based on our clinical observations (i.e., the ribbon sign and septum-intersection triangular sign), whose diagnostic potential warranted formal investigation. A total of 86 patients with pathologically confirmed lesions were included. Six key imaging variables with high diagnostic relevance were identified. Diagnostic models were subsequently constructed using either three or six selected features. Among the evaluated models, the XGBoost model demonstrated the highest diagnostic performance. Furthermore, a significant association was observed between intracystic hemorrhage and the presence of the ribbon sign.

Interpretation of findings

Within the study population, MHCs were more frequently observed as multiple lesions, whereas MCNs tended to be present as solitary lesions. This characteristic demonstrated both high sensitivity and specificity. Smooth internal septa were predominantly observed in MHCs, while irregular or nodular septa were more indicative of MCNs. The presence of nodular protrusions along the septa may reflect proliferative and invasive behavior of tumor cells (19). This feature exhibited high specificity but low sensitivity, indicative of its utility in disease exclusion rather than confirmation.

The ribbon sign, defined as a free-floating septum attached to the cyst wall at one end only, was more frequently observed in patients diagnosed with MHC. Although infrequently reported in the literature, the ribbon sign demonstrated high sensitivity but low specificity, indicating a greater likelihood of false-positive results. As such, it may serve as a useful screening marker but lacks sufficient precision for confirmatory diagnosis.

A cyst wall thickness of ≥2 mm was more frequently associated with MCNs. A plausible explanation is the presence of abundant stromal elements, including ovarian-type stroma, within the cyst walls of MCNs (20). In contrast, the cyst walls of benign non-neoplastic cysts, such as simple hepatic cysts, are typically thin and histologically less complex (21,22). This feature demonstrated high specificity but low sensitivity, reinforcing its role in the exclusion of MCN, rather than in its definitive identification.

We acknowledge the practical challenge in reliably discriminating a 2-mm from a 3-mm wall thickness on routine imaging. However, our standardized measurement protocol ensured internal consistency. The significant association we found, coupled with its exceptionally high specificity (0.967) in the model, suggests that the qualitative perception of definite wall thickening—as opposed to an ultrathin wall—serves as a powerful and practical imaging ‘red flag’ for raising suspicion of MCN and aiding in the exclusion of benign, thin-walled cysts.

The septum-intersection triangular sign was rarely observed, particularly in patients with MCN. This feature may result from localized increases in intracystic pressure or mechanical tension across fibrous septa. Currently, this feature has not been formally documented in published literature. Despite its rarity, this manifestation exhibited high specificity but low sensitivity, which supports its use in ruling out MCN.

Centrally located septa were more prevalent in MCNs, whereas peripheral septa were more characteristic of MHCs. Although the underlying mechanisms remain unclear, one hypothesis suggests that proliferative and secretory activity of the columnar epithelium lining the cyst wall and septa in MCNs may result in uneven intracystic pressure distribution, contributing to the formation of centrally positioned septa (4).

Among the four assessed machine learning models, the XGBoost model achieved the highest diagnostic accuracy regardless of whether three or six imaging features were incorporated. As an ensemble learning algorithm based on gradient boosting, XGBoost is well-suited for managing complex nonlinear relationships and demonstrates robustness in the presence of noise and outliers in the dataset (23). The SHAP method was used to quantify the contribution of each variable to the predictive classification (24). Lesion type exerted the most significant influence on the model’s predictions, followed by septum appearance and cyst wall thickness, findings consistent with those of previous studies (8,15,25). Although septum position demonstrated diagnostic relevance, its contribution was less consistent.

This study further assessed the diagnostic value of novel morphological variables. The pathological mechanisms underlying the solitary nature of MCNs have not yet been fully clarified. Notably, the ribbon sign and septum-intersection triangular sign were identified as significant predictive features in this analysis, findings that have not been previously emphasized in existing literature.

The diagnostic relevance of the ribbon sign in hepatic cystic lesions, along with its potential underlying mechanisms, was further examined in this study. Firth logistic regression analysis, adjusted for age and sex, demonstrated a significant association between the ribbon sign and intracystic hemorrhage. However, the significant association should be interpreted with consideration of complete data separation within the MCN subgroup. The exact magnitude of this effect within the MCN population remains uncertain due to the rarity of its co-occurrence. Importantly, no significant interaction was observed between this association and lesion type of disease type, suggesting that the relationship is independent of whether the lesion is classified as MHC or MCN. These findings imply that the observed correlation between the ribbon sign and intracystic hemorrhage may be more closely related to the pathological features and biomechanical properties of the cystic lesions, rather than lesion-specific histopathology. One possible explanation is that ribbon-like septa may possess relatively thin and flexible structural characteristics, rendering them more susceptible to mechanical disruption. Intracystic hemorrhage may exacerbate this vulnerability through increases in intracystic pressure or as a result of external mechanical stress.

The ‘ribbon sign’ and ‘septum-intersection triangular sign’, while not yet standard lexicon, were terms we adopted to describe recurrent and distinct imaging patterns observed in our cohort. Crucially, our data provide the first analytical evidence that the ribbon sign has a strong and independent association with intracystic hemorrhage, regardless of the underlying lesion type. This finding positions it not merely as a morphological curiosity but as a potential imaging biomarker hinting at specific biomechanical properties of septa that predispose to hemorrhage, bridging imaging morphology and pathophysiology.

The ultimate goal of this study is to bridge the gap between imaging findings and clinical management. Based on the quantified feature importance from our analysis, we propose a practical, hierarchical framework for evaluating MHC. In routine practice, the number and shape of the lesions should be evaluated first; a solitary lesion with irregular septa constitutes the strongest initial indicator for MCN. The presence of definite wall thickening (visually estimated as ≥2 mm) or centrally located septa should serve as secondary flags, significantly elevating the suspicion level. Furthermore, identification of the ribbon sign ought to trigger a search for intracystic hemorrhage and be considered supportive of MCN. For equivocal cases where features conflict or are atypical, the XGBoost model developed herein offers valuable assistance, integrating this multidimensional information to output an objective risk score that can inform individualized patient management strategies, such as intensifying follow-up or recommending surgical consultation.

Previous studies

Existing research on the radiologic differentiation between MHCs and MCNs remains relatively limited. Most prior studies have focused on distinguishing MCNs from simple hepatic cysts, with considerable variability in the imaging features examined across investigations. Several findings from the present study align with previously reported results.

A 2024 study employing an RF algorithm was used to identify four MRI features that differentiated MCNs from septated hepatic cysts. These features included septa originating from non-indented cyst walls, the presence of multiple septa, intracystic cysts, and solitary lesion configuration. The resulting model achieved an AUC of 0.982 (15). The current study similarly identified solitary lesion configuration as a discriminative feature. In multivariable analysis, the number of septa (≥5) approached but did not reach statistical significance (P=0.057) and was therefore not included in the final model, possibly due to variations in sample size, population heterogeneity, or the complexity of radiological features.

In another study published in 2021, machine learning was applied to differentiate MCNs from benign hepatic cysts. Key features included the presence of solid enhancing nodules, septa extending from external lobes, and lesion configuration (solitary vs. multiple). These indicators contributed to a classification accuracy of 93.5%, with results validated across multiple institutions (16,26).

A 2025 review of 577 patients reported that features such as wall thickness, the presence of septa or mural nodules, intracystic debris, wall enhancement, biliary dilation, and left-lobe lesion location were predictive of MCNs. In contrast, simple hepatic cysts were typically thin-walled, regular in shape, multiple in number, and located in the right hepatic lobe (8). In this study population, cyst wall thickness and septum appearance were identified as discriminative features. However, lesion location and biliary dilation were not found to be reliable indicators for distinguishing between MHC and MCN.

One possible explanation is that MHCs and MCNs typically lack direct communication with the biliary system (1,27). Biliary dilation is more often a secondary phenomenon caused by mass effect from large lesions and may therefore lack specificity in this diagnostic context (28). Although a statistically significant difference in lesion location was observed between MHCs and MCNs in this cohort, lesion location was selected only by the Boruta algorithm and not retained in the final models. This may be explained by limitations in sample size, or a more complex relationship between lesion location and disease type.

Differences were also noted regarding the relationship between the septum and cyst wall. In the present study, this variable did not demonstrate diagnostic significance. In both MHC and MCN cases, septa were observed originating from both indented and non-indented regions of the cyst wall within the same lesion. For the purposes of analysis, septa arising solely from the cyst wall and those originating from both the recess and the cyst wall, were categorized as arising from the cyst wall. This classification approach may have contributed to an overestimation of the number of septa categorized as originating from the cyst wall.

Study advantages and limitations

Unlike studies relying on a single statistical approach, our integration of three distinct feature-selection methods (Boruta, LASSO, and regression) mitigated selection bias. Subsequent model development with four algorithms and SHAP-based interpretability not only ensured robustness but also provided the first quantitative visualization of the relative contribution weights of key features in the MHC vs. MCN differential. This moves beyond identifying ‘useful features’ to understanding ‘how much each feature matters’ in the diagnostic decision process. Crucially, our data provide the first analytical evidence that the ribbon sign has a strong and independent association with intracystic hemorrhage.

However, certain limitations should be acknowledged. The relatively small sample size, particularly of the MCN group, may limit both the generalizability and predictive accuracy of the constructed models. Furthermore, external validation using independent datasets was not performed, thereby limiting the ability to assess model performance across different clinical settings. Another potential limitation is the subjective nature of interpreting certain radiological features, such as septum appearance and septum position, which may introduce interobserver variability. In addition, our study lacked systematic reports on enhanced-related data. Due to the retrospective design of this study, some included patients only underwent unenhanced CT or MRI examinations without contrast-enhanced scans, resulting in a certain degree of data deficiency regarding imaging features related to enhancement.

Conclusions

By integrating multi-modal feature selection and machine learning algorithms, this study developed a high-performance XGBoost diagnostic model. The model not only reaffirmed the paramount importance of known key features but also provided the first systematic quantification of their contribution weights. Furthermore, it presented evidence supporting the “ribbon sign” as a novel potential marker strongly linked to intracystic hemorrhage. For clinical practice, we propose a stratified diagnostic approach: during imaging review, priority should be given to lesion number (solitary lesions raising suspicion for MCN) and septal appearance (irregularity raising suspicion for MCN). The presence of cyst wall thickening (≥2 mm) or centrally located septa should significantly increase suspicion for MCN. The appearance of the “ribbon sign” should alert the radiologist to the possibility of intracystic hemorrhage and favor MCN. The combination of these imaging features offers a framework for more reliable, individualized, and non-invasive diagnosis of hepatic cystic lesions.

Acknowledgments

The authors would like to express their sincere gratitude to the radiologists and medical staff at Tianjin Medical University General Hospital and Tianjin Fifth Central Hospital for their invaluable contributions to data collection and imaging analysis. Special thanks also go to the patients who participated in this study for their willingness to contribute to medical research.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2595/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2595/dss

Funding: This study was supported by a grant from the Science and Technology Project of the Binhai New Area Health Commission of Tianjin (No. 2023BWKY004).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2025-1-2595/coif). The authors have no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the declaration of Helsinki. The study was approved by the Clinical Trial Ethics Committee of Tianjin Fifth Central Hospital (No. WZX-EC-KY2024031). This study is a retrospective analysis of medical records, and data anonymization is only used for undergraduate research. Therefore, the ethics committee waived the requirement for written informed consent.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Rawla P, Sunkara T, Muralidharan P, Raj JP. An updated review of cystic hepatic lesions. Clin Exp Hepatol 2019;5:22-9. [Crossref] [PubMed]
2. Bistritz L, Tamboli C, Bigam D, Bain VG. Polycystic liver disease: experience at a teaching hospital. Am J Gastroenterol 2005;100:2212-7. [Crossref] [PubMed]
3. Purdy AC, Grigorian A, Fernando D, Nahmias J, Demirjian AN. Intracystic Hemorrhage Complicating Polycystic Liver Disease in a 90-Year-Old Woman. ACG Case Rep J 2019;6:e00049. [Crossref] [PubMed]
4. Hutchens JA, Lopez KJ, Ceppa EP. Mucinous Cystic Neoplasms of the Liver: Epidemiology, Diagnosis, and Management. Hepat Med 2023;15:33-41. [Crossref] [PubMed]
5. Nambhanath S, Srinivas T, Bhat SP, Kailath MC, Babu SS. Mucinous CysticNeoplasm of the Liver: Case Report and Review of the Literature. Journal of Health and Allied Sciences NU 2025;15:142.
6. McIntyre CA, Girshman J, Goldman DA, Gonen M, Soares KC, Wei AC, Balachandran VP, Kingham TP, Drebin JA, Jarnagin WR, Gluskin JS, D'Angelica MI, Gerst SR. Differentiation of mucinous cysts and simple cysts of the liver using preoperative imaging. Abdom Radiol (NY) 2022;47:1333-40. [Crossref] [PubMed]
7. Yang J, Guo X, Zhang H, Zhang W, Song J, Xu H, Ma X. Differential diagnosis of pancreatic serous cystadenoma and mucinous cystadenoma: utility of textural features in combination with morphological characteristics. BMC Cancer 2019;19:1223. [Crossref] [PubMed]
8. Manzari Tavakoli G, Afsharzadeh M, Mobinikhaledi M, Behzad S, Ghorani H, Salahshour F. Differentiation between mucinous cystic neoplasms and simple cysts of the liver: a systematic review and meta-analysis. Abdom Radiol (NY) 2025;50:4125-38. [Crossref] [PubMed]
9. Lantinga MA, Gevers TJ, Drenth JP. Evaluation of hepatic cystic lesions. World J Gastroenterol 2013;19:3543-54. [Crossref] [PubMed]
10. Soni S, Pareek P, Narayan S, Varshney V. Mucinous cystic neoplasm of the liver (MCN-L): a rare presentation and review of the literature. Med Pharm Rep 2021;94:366-71. [Crossref] [PubMed]
11. Anderson MA, Dhami RS, Fadzen CM, Molina G, Taylor MS, Deshpande V, Qadan M, Catalano OA, Ferrone CR, Mojtahed A. CT and MRI features differentiating mucinous cystic neoplasms of the liver from pathologically simple cysts. Clin Imaging 2021;76:46-52. [Crossref] [PubMed]
12. Kim SU, Hwang JA, Han S, Lee JH, Choi SY, Ha SY. Refining imaging criteria for mucinous cystic neoplasm of the liver: simplified diagnostic approach. Eur Radiol 2025;35:4993-5003. [Crossref] [PubMed]
13. Furumaya A, Schulz HH, Verheij J, Takkenberg RB, Besselink MG, Kazemier G, Erdmann JI, van Delden OM. Diagnosis and treatment of patients with suspected mucinous cystic neoplasms of the liver: a retrospective cohort study. Langenbecks Arch Surg 2024;409:66. [Crossref] [PubMed]
14. Mavilia MG, Pakala T, Molina M, Wu GY. Differentiating Cystic Liver Lesions: A Review of Imaging Modalities, Diagnosis and Management. J Clin Transl Hepatol 2018;6:208-16. [Crossref] [PubMed]
15. Xiao SY, Xu JX, Shao YH, Yu RS. To identify important MRI features to differentiate hepatic mucinous cystic neoplasms from septated hepatic cysts based on random forest. Jpn J Radiol 2024;42:880-91. [Crossref] [PubMed]
16. Hardie AD, Chamberlin JH, Boyum JH, Sharbidre KG, Petrocelli R, Flemming BP, Zahid M, Venkatesh SK, Mruthyunjayappa S, Hajdu CH, Kovacs MD. Multi-Center Follow-up Study to Develop a Classification System Which Differentiates Mucinous Cystic Neoplasm of the Liver and Benign Hepatic Cyst Using Machine Learning. Acad Radiol 2022;29:1149-56. [Crossref] [PubMed]
17. Degenhardt F, Seifert S, Szymczak S. Evaluation of variable selection methods for random forests and omics data sets. Brief Bioinform 2019;20:492-503. [Crossref] [PubMed]
18. Ternès N, Rotolo F, Michiels S. Empirical extensions of the lasso penalty to reduce the false discovery rate in high-dimensional Cox regression models. Stat Med 2016;35:2561-73. [Crossref] [PubMed]
19. Huang J, Zhang L, Wan D, Zhou L, Zheng S, Lin S, Qiao Y. Extracellular matrix and its therapeutic potential for cancer treatment. Signal Transduct Target Ther 2021;6:153. [Crossref] [PubMed]
20. Khot R, Ganeshan D, Sundaram KM, Depetris JN, Ludwig DR. Cystic lesions and their mimics involving the intrahepatic bile ducts and peribiliary space: diagnosis, complications, and management. Abdom Radiol (NY) 2025;50:2969-88. [Crossref] [PubMed]
21. Morii K, Yamamoto T, Nakamura S, Okushin H. Infectious Hepatic Cyst: An Underestimated Complication. Intern Med 2018;57:2123-9. [Crossref] [PubMed]
22. Vachha B, Sun MR, Siewert B, Eisenberg RL. Cystic lesions of the liver. AJR Am J Roentgenol 2011;196:W355-66. [Crossref] [PubMed]
23. Zhou L, Zhu Q, Chen Q, Wang P, Huang H. Predicting hospital outpatient volume using XGBoost: a machine learning approach. Sci Rep 2025;15:17028. [Crossref] [PubMed]
24. Rezk NG, Alshathri S, Sayed A, El-Din Hemdan E, El-Behery H. XAI-Augmented Voting Ensemble Models for Heart Disease Prediction: A SHAP and LIME-Based Approach. Bioengineering (Basel) 2024;11:1016. [Crossref] [PubMed]
25. Boyum JH, Sheedy SP, Graham RP, Olson JT, Babcock AT, Bolan CW, Menias CO, Venkatesh SK. Hepatic Mucinous Cystic Neoplasm Versus Simple Biliary Cyst: Assessment of Distinguishing Imaging Features Using CT and MRI. AJR Am J Roentgenol 2021;216:403-11. [Crossref] [PubMed]
26. Furtado FS, Badenes-Romero Á, Hesami M, Mostafavi L, Najmi Z, Queiroz M, Mojtahed A, Anderson MA, Catalano OA. External validation of a machine learning based algorithm to differentiate hepatic mucinous cystic neoplasms from benign hepatic cysts. Abdom Radiol (NY) 2023;48:2311-20. [Crossref] [PubMed]
27. Lee MH, Katabathina VS, Lubner MG, Shah HU, Prasad SR, Matkowskyj KA, Pickhardt PJ. Mucin-producing Cystic Hepatobiliary Neoplasms: Updated Nomenclature and Clinical, Pathologic, and Imaging Features. Radiographics 2021;41:1592-610. [Crossref] [PubMed]
28. Lv Y, Liu N, Wu H, Li Z. Etiological classification and treatment strategies for secondary bile duct dilatation. Exp Biol Med (Maywood) 2021;246:281-5. [Crossref] [PubMed]