Artificial intelligence – assisted three-dimensional preoperative planning in liver tumor surgery: a comparative real-world study

Abbreviations:

• AI – artificial intelligence
• CT – computed tomography
• FNH – focal nodular hyperplasia
• TROPH-L – Tumor – Regional capsule – Outflow veins – Portal vein – Hepatic bile – Localization
• 3D – three-dimensional

Liver tumors represent a major global healthcare challenge and remain one of the leading causes of cancer-related mortality worldwide [1]. In routine clinical practice, modern tertiary care centers are increasingly required to make rapid, accurate, and evidence-based diagnostic and therapeutic decisions when managing patients with focal liver lesions [2]. This group of diseases remains particularly complex: radical treatment is often associated with a high risk of intraoperative and postoperative complications, while long-term outcomes are frequently suboptimal [3].

Importantly, comprehensive diagnostic assessment and surgical planning are required not only for malignant tumors but also for benign focal liver lesions. Large hemangiomas, adenomas, focal nodular hyperplasia (FNH), and complex cystic lesions may, in selected cases, cause significant pain, carry a risk of rupture, or lead to compression of vascular and biliary structures. As a result, these conditions can become clinically relevant and necessitate detailed preoperative evaluation [4].

In this context, there has been a growing interest in the integration of artificial intelligence (AI) technologies and three-dimensional (3D) modeling into clinical hepatobiliary surgery [5]. The combined use of AI-based tools and 3D visualization fundamentally transforms clinical decision-making by shifting it from subjective image interpretation toward objective, quantitative data analysis [6]. Contemporary deep learning algorithms have already demonstrated the ability to automatically segment hepatic structures on contrast-enhanced computed tomography (CT) images with an accuracy of up to 92–94%, improving diagnostic reproducibility while significantly reducing interpretation time [7].

The synergistic application of AI services and 3D modeling enables the creation of highly accurate, patient-specific virtual reconstructions of the liver, allowing detailed visualization of focal lesions and their anatomical relationships with vascular structures and the biliary tree [8]. This approach contributes to a more efficient diagnostic workflow, reducing the need for certain invasive procedures and decreasing the overall burden on the patient [9]. Clinical studies have shown that the use of 3D navigation and reconstruction techniques is associated with a significant reduction in operative time and a lower incidence of R1 resections, thereby directly improving the safety and oncological quality of surgical interventions [4][10]. Consequently, digital technologies not only streamline the diagnostic and planning stages but also enhance the overall quality of surgical care for patients with liver tumors [11].

Objective of the study: to evaluate the diagnostic and clinical effectiveness of the HepatoScan AI clinical decision support system which integrates neural network–based medical image analysis with interactive 3D visualization, in real-world clinical practice.

MATERIALS AND METHODS

Development of the neural network algorithm

The development and validation of the clinical decision support system HepatoScan AI for focal liver lesions was carried out in collaboration between Botkin Hospital and the Research and Practical Clinical Center for Diagnostics and Telemedicine Technologies. The system is officially registered in the Russian Federation as computer software [12].

A detailed description of the training process and the performance metrics for the classification of benign and malignant liver lesions has been published previously [13].

Development of the TROPH-L clinical and instrumental classification

To standardize the description of morphological and anatomical characteristics of focal liver lesions and to unify preoperative planning, we developed the TROPH-L classification (Tumor – Regional capsule – Outflow veins – Portal vein – Hepatic bile – Localization). This classification is designed for the systematic analysis of spatial relationships between the tumor and the major hepatic structures, based on data derived from 3D modeling and neural network–based segmentation.

The TROPH-L classification integrates morphological and topographic–anatomical features obtained from patient-specific virtual 3D models with clinically relevant resectability parameters. The classification criteria include:

• T (Tumor size and growth type) – lesion size and growth pattern.
• R (Relation to capsule) – relationship of the lesion to the hepatic capsule.
• O (Outflow veins involvement) – involvement of the hepatic veins.
• P (Portal vein involvement) – degree of involvement of portal vein branches.
• H (Hepatic bile involvement) – anatomical relationship with the biliary tree.
• L (Localization) – anatomical localization according to the Couinaud segmental system, specifying the lobe and segment.

Algorithm for three-dimensional preoperative modeling

For preoperative planning, a dedicated 3D modeling workflow was used, including the calculation of volumetric liver parameters and subsequent patient stratification according to the TROPH-L classification. Contrast-enhanced preoperative CT scans were imported into the 3D Slicer software environment in NIfTI format. At this stage, anatomical reconstruction of the liver was performed with semi-automatic segmentation of vascular and biliary structures [14].

In the second stage, the Slicer-Liver extension was applied to automatically delineate Couinaud liver segments and to calculate centerlines of the vascular trees. This was followed by virtual liver resection planning. Using the Resections module, the proposed resection plane was defined, with automatic calculation of hepatic parenchymal volumes, including total liver volume (TLV), resection volume (RV), remnant liver volume (RLV), and tumor volume (TuV).

The relative future liver remnant volume (aFLR, adjusted future liver remnant) was calculated automatically, reflecting the proportion of functionally preserved parenchyma and serving as a key safety criterion for the planned intervention (optimal threshold ≥ 30%). The finalized 3D models were exported in STL format, underwent post-processing in Blender software (Blender Online Community, version 4.0.2) with color labeling of anatomical structures, and were subsequently converted into glTF (.glb) format for use in 3D visualization and augmented reality systems¹.

The entire 3D modeling workflow was performed by surgeons from the specialized hepatopancreatobiliary surgery department with experience in 3D modeling, in close collaboration with a radiologist from Botkin Hospital. The initial automatic CT segmentation generated by the HepatoScan AI system was reviewed and, when necessary, manually corrected by the head of the radiology department (Maxim P. Onishchenko).

Final construction of the virtual liver models and determination of the planned RV were conducted through close interdisciplinary collaboration between the operating surgeons (Mikhail M. Tavobilov, Alexey A. Karpov, Aysa V. Lantsynova, Mark N. Aladin, Kirill A. Abramov, Evgeny B. Kudryash) and the 3D modeling specialist (Mark N. Aladin). The average time required to generate an individualized 3D liver model was approximately one hour (Fig. 1).

Patient stratification according to the TROPH-L classification

The proposed TROPH-L classification enables structured characterization of the morphologic and anatomical features of focal liver lesions and allows stratification of patients into three categories of surgical complexity, which is of critical importance for determining the extent of resection and selecting the optimal treatment strategy (Table 1).

Table 1. Surgical complexity categories according to the TROPH-L classification

Characteristic	I (low risk)	II (moderate risk)	III (high risk)
Lesion size and growth pattern	<5 cm, peripheral location	5–10 cm	centrally located or multifocal tumor
Relationship of the lesion to the hepatic capsule	subcapsular	subcapsular or intraparenchymal	subcapsular or intraparenchymal
Hepatic vein involvement	none	invasion of segmental veins	invasion of major hepatic vein branches
Portal vein branch involvement	none	invasion of segmental branches	invasion of major portal vein branches
Biliary hypertension	none	segmental or sectoral hypertension	lobal hypertension
Surgical planning	technically safe RV	use of 3D modeling to define the resection plane	detailed 3D reconstruction and assessment of the aFLR
Volumetric assessment	not required	calculation recommended: TLV, RV, RLV, TuV, aFLR	mandatory calculation: TLV, RV, RLV, TuV, aFLR
Segmentation example
Coding example	Tc1R1O0P0H0 – L4b	Ti4R0O1P2H1 – L7	Ti3R1O3P3H3R – L4–L7

Notes: green indicates the biliary ducts, purple represents the hepatic venous system, blue corresponds to the portal venous system, and yellow denotes the focal lesion.

aFLR – adjusted future liver remnant; RLV – remnant liver volume; RV – resection volume; TLV – total liver volume; TuV – tumor volume; 3D – three-dimensional.

This stratification makes it possible to anticipate the scope and technical complexity of surgery, the potential need for extended or combined resections, and the expected risk of postoperative complications. Integration of the TROPH-L classification with the HepatoScan AI system and 3D modeling algorithms provides a comprehensive preoperative planning framework based on objective anatomical and functional data. This approach ensures greater accuracy and consistency of surgical decision-making compared with traditional, largely subjective assessment methods.

Study design

A single-center comparative real-world study including a prospective AI-assisted cohort and a purposively matched retrospective cohort treated with conventional preoperative planning.

Sample size calculation

The required sample size was calculated based on the primary endpoint–differences in diagnostic performance, specifically sensitivity for detection and differential diagnosis of focal liver lesions. According to published data, the implementation of AI algorithms in radiological diagnostics is associated with an 8–12% increase in diagnostic sensitivity. Assuming an expected sensitivity improvement of at least 8%, a two-sided significance level of α = 0.05, and a study power of 80%, the minimum required sample size was estimated to be no fewer than 90 patients per group.

Patient enrollment

All patients who underwent surgical treatment in the specialized Hepatopancreatobiliary Surgery Department of Botkin Hospital between January 21, 2021, and June 27, 2025, were screened for eligibility.

Inclusion criteria:

• age ≥ 18 years;
• written informed consent for study participation;
• presence of a focal liver lesion requiring surgical treatment;
• availability of preoperative multiphasic contrast-enhanced abdominal CT;
• availability of intraoperative assessment data and/or histological verification.

Exclusion criteria:

• parasitic liver cysts;
• metastatic liver tumors and primary hepatic sarcomas;
• diagnostically non-informative CT images;
• incomplete clinical or radiological data.

The prospective AI-assisted cohort included patients treated between February 1, 2023, and June 27, 2025, in which the HepatoScan AI clinical decision support system integrated with 3D modeling was used during preoperative planning. Initially, 148 patients were assessed for inclusion; 44 patients met the exclusion criteria. A total of 104 patients were included in the final analysis (Fig. 2). The distribution by pathology was as follows: liver carcinomas (n = 30), including cholangiocarcinoma (CCA, n = 11) and hepatocellular carcinoma (HCC, n = 19), liver hemangiomas (n = 22), simple liver cysts (n = 30), and FNH (n = 22).

To minimize the impact of differences in pathological composition on diagnostic and surgical outcomes, a purposefully matched retrospective cohort was formed in accordance with established methodology for comparative surgical studies. Matching was performed at the subgroup level according to lesion type, with consideration of lesion size and key demographic characteristics.

For inclusion in the standard preoperative planning group (retrospective cohort), 174 patients with focal liver lesions who underwent surgical treatment between January 21, 2021, and January 25, 2023, were evaluated. Seventy-four patients were excluded based on the exclusion criteria. The final retrospective cohort consisted of 100 patients with the following diagnoses: liver carcinomas (n = 27), including CCA (n = 7) and HCC (n = 20), liver hemangiomas (n = 23), simple liver cysts (n = 30), and FNH (n = 20). Patient inclusion and allocation are illustrated in Figure 2.

Analysis of diagnostic effectiveness

All patients from both cohorts were included in the analysis of diagnostic effectiveness. For each patient, a unified preoperative diagnostic conclusion regarding the nature of the focal liver lesion was established and used to calculate diagnostic performance metrics. In patients with multiple lesions, analysis was performed at the level of the clinically dominant lesion that determined the surgical strategy.

Comparison of HepatoScan AI–assisted results with the reference standard was performed by expert assessment at the level of the final clinical diagnostic conclusion.

The reference standard varied depending on lesion type. For suspected malignant tumors, diagnostic accuracy was assessed by comparing the preoperative HepatoScan AI conclusion with results of preoperative morphological verification (biopsy), followed by confirmation of the final diagnosis based on histopathological examination of the resected specimen. For benign focal liver lesions, the reference standard was the conclusion of an experienced radiologist based on multiphasic contrast-enhanced CT, corroborated by intraoperative findings and the clinical course of the disease.

Analysis of clinical effectiveness

The analysis of clinical effectiveness of the HepatoScan AI clinical decision support system included patients from both cohorts with anatomically complex focal liver lesions corresponding to TROPH-L categories II–III.

Patients with TROPH-L category I lesions were excluded from this analysis: 50 patients from the prospective cohort and 45 patients from the retrospective cohort. All patients with simple liver cysts in both cohorts were excluded; all of them underwent cyst fenestration.

The final clinical effectiveness analysis included 54 patients from the prospective cohort and 55 patients from the retrospective cohort with liver carcinomas, hemangiomas, and FNH. The compared groups were evaluated with respect to operative time, intraoperative blood loss, and the rate of blood transfusions. In addition, oncological radicality in patients with liver carcinomas (rate of R0 resections), postoperative complication rates – including severe complications graded III–V according to the Clavien–Dindo classification – length of hospital stay, and in-hospital mortality were analyzed.

Surgical technique

All surgical procedures were performed using standard surgical techniques. The choice of surgical approach (open or laparoscopic) was determined on an individual basis, taking into account lesion localization, size, relationships with vascular and biliary structures, and the planned extent of resection. Liver parenchymal transection was carried out using an ultrasonic dissector and bipolar coagulation; in selected cases, an ultrasonic aspirator–destructor was employed with stepwise identification and control of vascular and biliary structures. The Pringle maneuver was applied selectively and only when indicated, in cases with an anticipated risk of significant intraoperative blood loss.

Statistical analysis

Statistical analysis was performed using IBM SPSS Statistics software version 30.0 (SPSS: An IBM Company, USA) and the R programming language (version 4.4.1), with the use of specialized packages pROC (version 1.18.0) and stats (base R package).

Quantitative variables were initially assessed for normality using the Shapiro–Wilk test. Depending on the distribution, data are presented as mean with standard deviation (M ± SD) or median with interquartile range (Me (Q1; Q3)). Between-group comparisons of continuous variables were performed using Student’s t-test for normally distributed data or the Mann–Whitney U test for non-normally distributed data.

Categorical variables are presented as absolute and relative frequencies and were compared using Pearson’s χ² test or Fisher’s exact test when expected cell counts were below five. Comparisons of blood transfusion rates between groups were performed using aggregated data and a Poisson regression model to calculate incidence rate ratios (IRR).

Diagnostic performance of the HepatoScan AI system was evaluated on a per-patient basis and included calculation of sensitivity, specificity, and the area under the ROC curve (AUROC), with corresponding 95% confidence intervals (CI).

For malignant focal liver lesions, the reference standard was based on preoperative morphological verification (biopsy) followed by confirmation of the final diagnosis through histopathological examination of the resected specimen. For benign focal liver lesions, the reference standard consisted of the radiologist’s conclusion based on multiphasic contrast-enhanced CT, corroborated by intraoperative findings and the clinical course of the disease. Comparisons of AUROC values between groups were performed using the nonparametric DeLong test.

Differences were considered statistically significant at p < 0.05.

RESULTS

Diagnostic performance of the HepatoScan AI system

The study groups did not differ significantly with respect to key demographic characteristics, including age, sex, and the presence of comorbid or background diseases. No statistically significant differences were observed between groups in terms of lesion size, number, multiplicity, or anatomical localization. Retrospective assessment of case complexity in the standard preoperative planning group using the TROPH-L classification demonstrated a distribution of risk categories comparable to that of the prospective AI-assisted preoperative planning group (Table 2).

When individual nosological subgroups were analyzed, no statistically significant differences in sensitivity, specificity, or AUROC of the ROC curve were observed between the groups. Across all subgroups, a consistent trend toward higher diagnostic performance metrics was noted in the AI-assisted preoperative planning group; however, due to the limited size of the subgroups, these differences did not reach statistical significance. When diagnostic metrics were compared across all types of focal liver lesions, integration of the HepatoScan AI system was associated with improved diagnostic sensitivity for focal liver lesions compared with the standard diagnostic approach (Table 3).

Clinical effectiveness of the HepatoScan AI system and three-dimensional preoperative modeling

The use of 3D modeling was most clinically justified in patients classified as TROPH-L categories II and III. This subgroup included 54 patients from the prospective cohort and 55 patients from the retrospective cohort. The compared groups did not differ significantly in terms of key clinical and demographic characteristics, prevalence of comorbidities, lesion size, anatomical localization, lesion multiplicity, or – among patients with carcinomas – tumor stage and histological grade (Table 4).

At the surgical stage, integration of the neural network–based clinical decision support system HepatoScan AI in combination with 3D preoperative modeling had a meaningful impact on the choice of surgical strategy in patients classified as TROPH-L categories II–III. In the AI-assisted preoperative planning group, a trend toward a more parenchyma-sparing approach was observed: extended anatomical liver resections were performed less frequently (27.8% vs 38.2% in the standard preoperative planning group), while the proportion of non-anatomical (atypical) resections was higher (55.6% vs 45.5%). The frequency of segmentectomies did not differ significantly between groups (Table 5).

In the subgroup of patients with carcinomas, R0 resection was achieved in all patients in the AI-assisted preoperative planning group (100%) compared with 91.3% in the standard preoperative planning group; however, this difference did not reach statistical significance (p = 0.18).

The use of AI technologies and 3D modeling was associated with improved intraoperative parameters. In the AI-assisted preoperative planning group, the mean operative time was reduced by approximately 20 minutes. A more pronounced effect was observed for intraoperative blood loss: mean blood loss in the AI-assisted group was 120 mL lower than in the standard preoperative planning group. In the AI-assisted preoperative planning group, a total of 4 units of packed red blood cells were transfused in 54 patients, corresponding to 0.07 units per patient, whereas in the standard preoperative planning group 13 units were transfused in 55 patients (0.24 units per patient). Accordingly, the rate of blood transfusion use was approximately 3.2 times higher in the standard preoperative planning group (incidence rate ratio 3.19; 95% CI 1.04–9.76).

The reduction in intraoperative surgical trauma was reflected in postoperative outcomes. In the AI-assisted preoperative planning group, a trend toward a lower overall rate of postoperative complications (28% vs 40%) as well as severe complications (Clavien–Dindo grades III–V: 9% vs 18%) was observed compared with the standard preoperative planning group; however, given the sample size, these differences did not reach statistical significance (p = 0.17).

In the AI-assisted preoperative planning group, severe complications were recorded in five patients and included postoperative bleeding (n = 3), pulmonary embolism (n = 1), and a biliary fistula Grade B according to the International Study Group of Liver Surgery (ISGLS) classification (n = 1) requiring percutaneous drainage under ultrasound guidance. No cases of acute postoperative liver failure were observed in this group. In the standard preoperative planning group, severe postoperative complications (n = 10) included acute postoperative liver failure (n = 3), clinically significant postoperative bleeding (n = 4), pulmonary embolism (n = 1), and biliary complications in the form of ISGLS Grade B biliary fistulas (n = 2), which required percutaneous drainage under ultrasound guidance.

In-hospital mortality was observed only in the standard preoperative planning group (n = 2); in both cases, death was attributable to progression of acute postoperative liver failure following extensive liver resections.

DISCUSSION

The results of the present study demonstrate that integration of AI technologies into the diagnostic workflow for liver tumors is associated with improved detection of focal liver lesions. Compared with the standard diagnostic approach, use of the HepatoScan AI system resulted in a statistically significant increase in the sensitivity of preoperative diagnosis for all focal liver lesions by 9.1% (from 84.2% to 93.3%), while maintaining a high level of specificity (91.8%).

The observed improvement in diagnostic accuracy with AI assistance is consistent with current literature. Several studies have shown that deep learning models are capable of detecting hepatic tumor lesions with accuracy comparable to expert radiologist assessment and, in some cases, identifying additional lesions not recognized during conventional visual analysis [6][9]. Moreover, it has been reported that the combined use of expert interpretation and AI algorithms increases the sensitivity of focal liver lesion detection from approximately 80% to 88% [15].

Visualization of the liver as an interactive 3D model enables “virtual navigation” within the organ, precise calculation of RV, and prediction of the risk of postoperative liver failure – features that are particularly important in cases of complex tumor localization. The TROPH-L classification developed and first introduced in this study integrates AI-based segmentation with 3D reconstruction and illustrates the feasibility of using objective morphological and topographic–anatomical parameters for standardized stratification of surgical risk. Overall, this approach aligns with the global trend toward personalized surgical treatment based on large-scale data analysis and the application of AI [16].

In the present study, automated 3D reconstruction of liver anatomy demonstrated clear clinical benefits in patients with moderate and high surgical risk, corresponding to TROPH-L categories II–III. Compared with standard preoperative planning, the AI-assisted group showed more favorable intraoperative and postoperative outcomes: operative time was reduced by an average of 11%, intraoperative blood loss was approximately 30% lower, and the need for blood transfusions was reduced by about 3.2-fold. These findings reflect improved selection of surgical approach and extent of resection based on detailed preoperative 3D anatomical analysis.

Our results are consistent with data from multicenter studies demonstrating that the use of 3D visualization in liver resection planning leads to statistically significant reductions in operative time and blood loss, lower rates of postoperative liver failure and other complications, and shorter hospital stays [10][11]. In particular, a recent prospective study showed that preoperative assessment using 3D reconstruction is associated with reduced surgical trauma and faster postoperative recovery compared with standard two-dimensional navigation [11].

The impact of 3D planning on oncological outcomes should also be emphasized. According to published data, the use of 3D navigation facilitates more accurate definition of resection margins and increases the rate of R0 resections [17]. In the present study, no cases of positive surgical margins (R1 resections) were observed in the group undergoing 3D reconstruction, whereas such cases occurred, albeit infrequently, in the standard preoperative planning group (2 of 23 patients, 8.7%). These findings are in agreement with meta-analyses demonstrating that 3D-assisted liver resections are associated with wider resection margins and lower recurrence rates [11][16][17].

Furthermore, more accurate calculation of the aFLR and detailed 3D visualization of tumor relationships with vascular and biliary structures may have contributed to a higher rate of parenchyma-sparing yet oncologically radical resections, avoiding unjustifiably extensive liver resections. This approach was associated with a trend toward a lower incidence of severe postoperative complications (Clavien–Dindo grade ≥ III) and the absence of perioperative mortality, reflecting improved surgical safety [17][18].

Taken together, these findings suggest that integration of neural network–based algorithms and 3D modeling can serve as an effective decision-support tool for personalized surgical planning. Such an approach has the potential to enhance the quality of preoperative diagnosis and improve short-term surgical outcomes in patients with focal liver lesions, making operative planning more objective and reproducible and surgical interventions safer and more effective.

At the same time, widespread implementation of these systems requires further investigation and resolution of several organizational and technical challenges. It should be acknowledged that most studies on AI-based liver imaging are based on retrospective data from relatively small cohorts [19], which limits the strength of conclusions. Nevertheless, the rapid advancement of machine learning technologies and the accumulation of prospective clinical data provide grounds to anticipate an increasingly prominent role for such tools in routine clinical practice.

Limitations

When interpreting the results of this study, several limitations should be considered. First, the comparison between a retrospective cohort (2021–2023) and a prospective cohort (2023–2025) may reflect not only the impact of digital technology implementation but also temporal changes in surgical techniques and perioperative patient management. Second, the relatively small sample size – particularly after stratification by nosological subgroups and TROPH-L complexity categories – limits the statistical power of the analysis and the ability to detect differences in certain clinical outcomes. In addition, different reference standards were applied for different types of focal liver lesions, and comparisons involving the HepatoScan AI system were performed at the level of the final clinical diagnostic conclusion, which does not fully exclude the influence of expert interpretation. Finally, the single-center design of the study and the use of the proprietary TROPH-L classification may limit the generalizability of the findings and their direct comparability with results from other studies.

CONCLUSION

In this single-center comparative study, integration of neural network–based CT analysis with interactive 3D preoperative modeling (HepatoScan AI) was associated with increased sensitivity for detection and classification of focal liver lesions compared with the standard approach. Automated segmentation, quantitative 3D volumetry, and the TROPH-L classification enabled standardized assessment of anatomical complexity and resectability, which was particularly relevant for patients classified as TROPH-L categories II–III. In this subgroup, AI assistance and 3D planning were associated with improvements in selected intraoperative parameters and a trend toward a lower incidence of severe postoperative complications, while maintaining oncological radicality. Further prospective multicenter studies are warranted to standardize methodologies for evaluating digital tools in liver surgery and to validate these findings across broader clinical settings.

AUTHORS CONTRIBUTIONS

Alexey V. Shabunin and Mikhail M. Tavobilov conceptualized and designed the study and contributed to the surgical procedures. Mark N. Aladin curated the clinical data, contributed to the surgical procedures, performed 3D image segmentation and post-processing, conducted data analysis, and contributed to the development and validation of artificial intelligence algorithms and 3D modeling. Alexey A. Karpov contributed to the surgical procedures and provided organizational and methodological oversight of the surgical component of the study. Aysa V. Lantsynova and Kirill A. Abramov contributed to the surgical procedures, conducted the literature review, and supported clinical data analysis. Evgeny B. Kudryash contributed to the surgical procedures and performed statistical analysis and interpretation of the results. Alexey V. Shabunin, Mikhail M. Tavobilov, and Mark N. Aladin contributed to manuscript drafting and critical revision. All authors reviewed and approved the final manuscript.

Ethics statements. The study was approved by the Local Ethics Committee (Protocol No. 15 dated 11 October 2022) of the Russian Medical Academy of Continuous Professional Education, Ministry of Health of the Russian Federation.

Data availability. The data confirming the findings of this study are available from the authors upon reasonable request. Data and statistical methods used in the article were examined by a professional biostatistician on the Sechenov Medical Journal editorial staff.

Conflict of interest. The authors declare that there is no conflict of interests.

Financing. This study was supported by a government-funded Research and Development project aimed at reducing postoperative complications and mortality and improving survival outcomes in patients with benign and malignant liver diseases through the use of neural network technologies and preoperative three-dimensional modeling (EGISU No. 123042600026-2). The project was approved by the Moscow City Health Department under Order No. 1196 dated December 21, 2022, covering the period 2023–2025.

Acknowledgements. The authors express their deep gratitude to the Head of the Department of Radiology of Botkin Hospital Maxim P. Onishchenko for his assistance in organizing radiological examinations, as well as to the staff of the Research and Practical Clinical Center for Diagnostics and Telemedicine Technologies: Chief Physician Yuri A. Vasiliev, Head of the Directorate for Scientific Division Management Olga V. Omelyanskaya, Head of the Department of Medical Informatics, Radiomics, and Radiogenomics Kirill M. Arzamasov, Deputy Head of the Department of Medical Informatics, Radiomics and Radiogenomics Lev D. Pestrenin, and Junior Researcher of the Department of Radiomics and Radiogenomics Ekaterina F. Savkina, for their contribution to the development of the HepatoScan AI system and to the processing and analysis of radiological examinations.