A biology-informed radiomics model for prognostication of hepatocellular carcinoma based on AKR1B10 expression.

1
Introduction
Hepatocellular carcinoma (HCC) represents a major global health challenge, ranking as the third leading cause of cancer-related mortality worldwide [1]. Despite advances in treatment modalities, the 5-year overall survival rate remains only 18 %, with postoperative recurrence rates reaching 70 % within five years [2], [3].
Current prognostic indicators, including alpha-fetoprotein (AFP) and protein induced by vitamin K absence or antagonist-II (PIVKA-II), lack the sensitivity and specificity required for precise risk stratification [4]. This clinical reality underscores the urgent need for novel biomarkers that can accurately predict tumor behavior and guide personalized treatment strategies.
Radiomics has emerged as a promising approach for HCC prognostication, with numerous studies demonstrating successful prediction of survival, recurrence, and microvascular invasion [5], [6]. However, these models predominantly rely on direct correlations between imaging features and clinical outcomes, creating a "black box" phenomenon that lacks biological interpretability [7]. This disconnect between prediction and biological understanding limits clinical translation, as physicians require mechanistically grounded tools to inform treatment decisions. Therefore, developing biology-informed radiomic models that link imaging phenotypes to specific molecular drivers represents a critical advancement toward clinically meaningful prognostic tools.
In recent years, more and more studies have demonstrated that AKR1B10 can be utilized as a diagnostic marker for HCC, with promising results in the detection of early-stage HCC and AFP-negative patients [8]. AKR1B10 encodes a member of the aldehyde/ketone reductase superfamily, which includes more than 40 recognized enzymes and proteins. RNAseq and Microarray both revealed the gene expression of AKR1B10 in the liver. A large-scale multicenter study validates AKR1B10 as a prevalent marker for the detection of HCC, combining AKR1B10 and AFP enhanced HCC diagnostic accuracy when compared to AKR1B10 or AFP alone [9]. Moreover, Wang et al. demonstrated that the suppression of AKR1B10 expression resulted in cell cycle arrest and inhibited cell proliferation, suggesting its potential tumorigenic role in facilitating HCC cell growth [10].
This study aims to develop a biology-informed radiomics model that predicts AKR1B10 expression status, thereby bridging imaging phenotypes with underlying tumor biology. Rather than directly correlating images with survival, a complex endpoint influenced by multiple confounders, we hypothesize that predicting a specific molecular driver provides both prognostic value and biological interpretability. Our approach addresses three clinical needs: (1) enabling non-invasive assessment of AKR1B10 status without tissue sampling, (2) providing biological context for risk stratification beyond conventional "black box" predictions, and (3) identifying candidates for emerging AKR1B10-targeted therapies [11]. By integrating this radiomics signature with clinical variables, we aim to establish a comprehensive prognostic tool that advances precision medicine in HCC.

2
Materials and methods
2.1
Data source and study cohort
Clinical and Genomic Cohort This retrospective study utilized data from TCGA. We initially identified 377 patients with hepatocellular carcinoma (LIHC). The inclusion criteria: (1) diagnosis of primary HCC confirmed by pathology; and (2) surgical resection performed as the initial curative treatment. And the exclusion criteria: (1) patients with unknown survival status or missing follow up data (n = 42); (2) patients with an overall survival of less than 30 days were excluded to eliminate the confounding effect of perioperative mortality (n = 18); and (3) patients with incomplete clinical annotations required for multivariate analysis (missing RNA-seq and TNM stage data) (n = 22). Following this selection process, 295 patients were included in the final clinical-genomic cohort (Supplementary Figure 1).
Radiogenomic Cohort For the development of the radiomics model, we sought matched imaging data from The Cancer Imaging Archive (TCIA). As a public service hosted by the University of Arkansas for Medical Sciences, TCIA provides de-identified, DICOM-format medical images. We retrieved corresponding contrast-enhanced CT images for the filtered TCGA cohort. Among the 295 eligible patients, 70 had available images in the TCIA-LIHC collection. Thirty-six patients were subsequently excluded due to poor image quality. Thus, the final radiogenomic cohort consisted of 34 patients with complete matched RNA-seq, clinical, and imaging data (Supplementary Figure 2).

2.2
Survival and enrichment analysis of AKR1B10
All survival and enrichment analyses were performed on the TCGA cohort (n = 295). The optimal cutoff value for stratifying patients into AKR1B10 high- and low-expression groups was determined using the “surv_cutpoint” function of the survminer R package [12]. Univariate and multivariate Cox proportional hazards regression models were employed to estimate hazard ratios (HRs) and 95 % confidence intervals (CIs) for overall survival. For the univariate and multivariate Cox proportional hazards regression analyses, the following variables were entered: AKR1B10 expression status, age, sex, pathologic stage, vascular invasion, histologic grade, hepatic inflammation, ablation/embolization history, pharmaceutical therapy, AFP level, and residual tumor status. Variables with P < 0.1 in univariate analysis were included in the multivariate model. The association between AKR1B10 expression and clinicopathological features was assessed using Pearson’s Chi-squared test or Fisher’s exact test, as appropriate. Differences in immune cell infiltration between the two expression groups were analyzed using the CIBERSORTx algorithm and compared via the limma R package. Gene Set Enrichment Analysis (GSEA) was conducted using the clusterProfiler R package to identify Hallmark and KEGG pathways significantly enriched in the AKR1B10 high-expression group.

2.3
Tumor segmentation and radiomics feature extraction
Tumor segmentation was performed using a semi-automated approach with 3D Slicer software (version 4.11.0). Initial segmentation was generated using the "GrowCut" algorithm with a neighborhood radius of 3 voxels and strength coefficient of 0.2. Two board-certified radiologists (5 and 7 years of experience) independently refined segmentations to ensure accurate tumor boundary delineation. For each lesion, a 3D volume of interest (VOI) encompassing the entire tumor was defined on portal venous phase CT images, with a 3-mm peritumoral margin generated using SimpleITK Python package [13], [14]. Inter-observer agreement was assessed on 10 randomly selected cases, with intraclass correlation coefficients (ICC) ranging from 0.85 to 0.94 for key radiomic features. In cases with > 10 % volumetric discrepancy between readers, consensus review was conducted with a senior radiologist (12 years of experience). All images were resampled to isotropic voxel size (1 ×1 ×1 mm³) to ensure spatial uniformity. Using PyRadiomics (version 3.0), we extracted 107 features including: 18 first-order statistical, 14 shape-based, and 75 texture features (24 GLCM, 16 GLRLM, 16 GLSZM, 14 GLDM, 5 NGTDM). Features were normalized using z-score transformation (Fig. 1).

2.4
Development of AKR1B10-informed radiomics model
We developed two separate deep learning models to investigate the prognostic significance of the peritumoral microenvironment: The Whole tumor Model (T Model) took cropped images containing only the tumor region as input. The Whole and Peri-tumor Model (T + P Model) took images containing both the tumor and the expanded peritumoral region (dilated by 3 mm) as input, allowing the network to learn patterns from both the tumor core and the invasive margin. Feature selection employed a two-step process: maximum-relevance minimum-redundancy (mRMR) algorithm selected 20 features based on strong outcome association and low inter-feature correlation, followed by recursive feature elimination (RFE) to refine the feature set. The final subset was used to construct a logistic regression model outputting a radiomics score (Rad-score) representing probability of high AKR1B10 expression. Model performance was evaluated through ROC and precision-recall (PR) curves, with metrics including area under the ROC curve (AUC), Brier score, accuracy, sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Calibration was assessed using calibration curves and Hosmer-Lemeshow test. Clinical utility was evaluated via decision curve analysis (DCA). Internal validation used 5-fold cross-validation. Differences in Rad-scores between AKR1B10 high/low groups were compared using Wilcoxon rank-sum test.

3
Statistical analysis
Numerical data are presented as mean ± standard deviation, and categorical data as relative frequencies and percentages. Chi-square tests compared categorical baseline characteristics between groups. Kaplan-Meier curves and log-rank tests compared survival outcomes. Cox proportional hazards regression performed survival analysis.
For subgroup analysis, data were stratified to assess interactions between AKR1B10 expression and covariates. Spearman's correlation coefficients measured associations between AKR1B10 and other variables. Statistical significance was set at a two-sided P < 0.05. Analyses were conducted using R v4.2.1 (https://www.r-project.org).

4
Results
4.1
Biological relevance of AKR1B10 in HCC
We conducted a comprehensive analysis of 295 patients from the TCGA database, stratifying them into AKR1B10 high expression (n = 148) and low expression (n = 147) groups using a cutoff value of 6.779. The baseline clinical characteristics between these two groups were well-balanced, with no statistically significant differences observed except for gender distribution (p > 0.05, Table 1). In the radiogenomic cohort of 34 patients with available imaging data (Table 2), AKR1B10 expression status was associated with several clinically relevant parameters. High AKR1B10 expression was significantly associated with elevated AFP levels (P = 0.040) and the presence of vascular invasion (P = 0.049). Patients with AKR1B10-high tumors demonstrated significantly reduced median overall survival compared with those with low expression (18.6 months vs. 31.4 months, P = 0.022). Furthermore, the AKR1B10-high group exhibited more aggressive disease characteristics, including higher recurrence rates (60.0 % vs. 31.6 %) and significantly shorter median time to recurrence (12.3 months vs. 24.8 months, P = 0.031).

4.2
Prognostic significance of AKR1B10 in HCC
AKR1B10 was significantly overexpressed in HCC tissues compared with 160 normal liver tissues from GTEx (median difference: 4.41-fold, P < 0.001). High AKR1B10 expression emerged as an independent predictor of poor survival in both univariate (HR = 1.811, 95 % CI: 1.202–2.730, P = 0.005) and multivariate analyses (HR = 2.187, 95 % CI: 1.385–3.454, P < 0.001). Advanced pathological stage (III/IV) similarly predicted poor outcomes in both models. Subgroup analysis revealed no significant interactions between AKR1B10 and clinical variables (Fig. 2A-C).
This overexpression carried substantial prognostic significance: patients with high AKR1B10 expression demonstrated markedly reduced overall survival compared with those with low expression (median OS: 52 vs. 84.4 months, P = 0.004; Fig. 2D). AKR1B10 expression showed associations with gender (P = 0.011) and pathological stage (P = 0.042; Fig. 2E).

4.3
Molecular and immune landscape
GSEA identified enrichment of oncogenic pathways in AKR1B10-high tumors, including TNFA_SIGNALING_VIA_NFKB, KRAS_SIGNALING_UP, and MTORC1_SIGNALING in the Hallmark gene set (Fig. 2F), along with immune related pathways such as NOD_LIKE_RECEPTOR and TOLL_LIKE_RECEPTOR signaling in KEGG analysis (Fig. 2G). The AKR1B10-high group exhibited distinct immune infiltration patterns, with significantly increased regulatory T cells, M0 macrophages, and resting dendritic cells (P < 0.001), while CD8 + T cells and B cells showed no significant differences (Fig. 2H).

4.4
Construction and evaluation of the radiomics model
From the intersection of TCGA and TCIA datasets, 34 patients with matched data were analyzed. A total of 107 radiomic features were extracted using PyRadiomics (v3.0). Feature selection through mRMR and recursive feature elimination identified 3 optimal features for the T model (Whole_shape_Sphericity, Whole_shape_Elongation, and Whole_glszm_Small Area Emphasis; Figs. 3A-B) and 4 features for the T + P model (Whole_peri_ngtdm_Busyness, Whole_peri_shape_Flatness, Whole_peri_ shape_Sphericity, and Whole_peri_glcm_ClusterShade; Fig. 4A-B).

4.5
Model performance
Both models demonstrated robust predictive performance for AKR1B10 expression status (Table 3). The T model achieved an AUC of 0.78 (95 % CI: 0.61–0.94), maintaining stability with cross-validation AUC of 0.76 (95 % CI: 0.58–0.94; Fig. 3C-E). The T + P model showed superior discrimination with an AUC of 0.83 (95 % CI: 0.69–0.97) and cross-validation AUC of 0.80 (95 % CI: 0.66–0.95; Fig. 4C-E). Calibration analysis confirmed excellent model fit (Hosmer-Lemeshow test: P = 0.88 and P = 0.44, respectively; Figs. 3F, 4F). Decision curve analysis validated clinical utility for both models (Figs. 3G, 4G).
Importantly, Rad_scores were significantly higher in the AKR1B10-high expression group for both models (T model: P = 0.006; T + P model: P < 0.001; Figs. 3H, 4H). While DeLong testing showed no significant difference between models (P = 0.447 for training, P = 0.695 for cross-validation; Fig. 5A-B), the T + P model was selected based on its slightly superior AUC. Feature reliability was excellent, with ICC values of 0.99 and 0.98 for the respective models, and all selected features demonstrating ICC > 0.85 (Fig. 5C).

4.6
Construction and validation of the T + P model based prognostic nomogram
Based on the multivariate Cox regression analysis, we identified Radiomics Score, Gender, and Hepatic Inflammation as independent prognostic factors. These variables were integrated to construct a visual prognostic nomogram for predicting the overall survival (OS) of HCC patients (Fig. 6A). In this nomogram, each variable is assigned a score on the "Points" scale based on its value. By summing the scores of all variables to obtain the "Total Points," the predicted median survival time and the probabilities of 1-, 3-, and 5-year survival can be estimated for individual patients via the linear predictor.
The predictive performance of the nomogram was evaluated using time-dependent ROC analysis. As illustrated in Fig. 6B, the model demonstrated satisfactory discrimination, with Area Under the Curve (AUC) values of 0.69, 0.85, and 0.82 for predicting 1-, 3-, and 5-year survival, respectively. The model showed particularly robust performance in predicting mid-to-long-term survival (3 and 5 years). Additionally, the C-index analysis (Fig. 6C) corroborated the model's discrimination ability between the predicted and observed outcomes. The dynamic AUC distribution over time (Fig. 6D) further confirmed the stability of the model's predictive accuracy throughout the observation period.
To assess the clinical translatability of the model, Decision Curve Analysis (DCA) was performed. The DCA curves (Fig. 6E) indicated that using the nomogram to predict 1-, 3-, and 5-year survival provided a greater net benefit across a wide range of threshold probabilities compared to the "treat-all" or "treat-none" strategies. This suggests that the radiomics-based nomogram holds significant potential for guiding clinical decision-making and personalized prognosis assessment.

5
Discussion
In this study, we successfully developed and validated a biology-informed radiomics model designed to non-invasively predict AKR1B10 expression status in HCC. Our principal finding is that integrating intratumoral and peritumoral radiomic features yields a robust imaging signature capable of accurately stratifying patients based on AKR1B10 expression levels. Notably, the inclusion of peritumoral features improved model performance compared to intratumoral features alone, suggesting that the biological influence of AKR1B10 extends beyond the visible tumor boundaries. By linking these macroscopic imaging phenotypes to specific oncogenic pathways (such as NF-κB and mTORC1 signaling) and immune infiltration patterns, our model bridges the gap between radiomics and tumor biology. Furthermore, the integration of this radiomics signature into a clinical nomogram demonstrated high accuracy for predicting patient survival, thereby providing a clinically applicable and biologically interpretable tool for pre-operative risk assessment.
The prominence of peritumoral features in our predictive model offers a macroscopic window into the biological behavior of AKR1B10 during HCC progression. Biologically, AKR1B10 plays a critical role in lipid metabolism and carbonyl detoxification. These metabolic activities likely induce microenvironmental remodeling that extends into the peritumoral hepatic parenchyma, establishing a pro-tumorigenic niche characterized by altered cellular density and metabolic gradients. This biological heterogeneity manifests as detectable variations in CT attenuation, which are captured by the specific texture features identified in our study, such as Whole_peri_ngtdm_Busyness and Whole_peri_glcm_ClusterShade. Furthermore, the selection of shape-based features (e.g., Sphericity) suggests that AKR1B10-mediated metabolic reprogramming may influence tumor growth patterns and boundary invasiveness. By correlating these specific imaging phenotypes with a validated molecular driver, our approach moves beyond the "black box" limitations of conventional radiomics, providing a biologically grounded explanation for the model's prognostic capability.
Our AKR1B10-based radiomics model achieved an AUC of 0.83 in predicting molecular status. To validate the feasibility of this biology-driven approach, we compared our results with established radiogenomic models targeting other HCC biomarkers. Our model demonstrated performance comparable to or exceeding that of models predicting VEGF (AUC 0.74) [15], Ki-67 (AUC 0.79) [16], and PD-L1 (AUC 0.76) [13]. It confirms that AKR1B10 expression translates into a distinct and detectable radiophenotype on CT, validating its utility as a robust imaging biomarker. Unlike 'black box' radiomics models that directly correlate texture with survival, our high predictive accuracy for AKR1B10 ensures that the subsequent prognostic stratification is grounded in a verified molecular mechanism. The superior detectability of AKR1B10 compared to some other markers suggests that its metabolic impact on the tumor microenvironment creates a particularly strong imaging signature, making it an ideal candidate for biology-informed prognostication.
The clinical relevance of our radiomics model is further underscored by the relationship between AKR1B10 and serum AFP levels. In our cohort, we observed no significant correlation between AKR1B10 expression and AFP, suggesting that these two biomarkers reflect distinct biological pathways in HCC tumorigenesis. This independence is crucial for the clinical utility of our proposed model. While AFP remains the standard prognostic marker, it is non-diagnostic in up to 40 % of patients (AFP-negative HCC) [17]. Our findings indicate that the AKR1B10-predicted radiomics score can effectively stratify risk regardless of AFP status. Therefore, our model does not merely replicate the information provided by serum AFP but serves as a complementary imaging biomarker, particularly valuable for refining prognosis in AFP-negative patients who currently lack reliable risk stratification tools.
Our feature selection strategy combining mRMR and RFE algorithms effectively identified biologically relevant imaging signatures while maintaining model interpretability. The selection of only 3–4 features prevents overfitting in our limited sample size while ensuring clinical feasibility. Unlike deep learning approaches that may achieve marginally higher accuracy but function as "black boxes," our feature-based method allows direct interpretation of which imaging characteristics associate with AKR1B10 expression. The excellent inter-reader agreement (ICC > 0.85 for selected features) confirms the reproducibility of our approach, essential for clinical implementation.
While our study focused on CT imaging due to its widespread availability and standardization in HCC management, incorporating multiparametric Magnetic Resonance Imaging (MRI) could potentially enhance model performance. MRI's superior soft tissue contrast and functional sequences ((Diffusion-Weighted Imaging (DWI), Dynamic Contrast-Enhanced (DCE)) might capture additional aspects of AKR1B10-related metabolic alterations. However, CT-based models offer practical advantages including lower cost, shorter acquisition time, and broader accessibility, particularly in resource-limited settings where molecular testing is unavailable. Future studies should explore multimodal imaging integration to determine whether the added complexity justifies potential performance gains.

6
Limitations
Although our study provides insights into the relationship between radiomics and AKR1B10 expression, several important limitations remain:
First, this study employed a retrospective single-center design with a relatively limited sample size, which may restrict the generalizability of the model. While we implemented internal validation, the absence of external validation cohorts from different institutions utilizing different CT scanning parameters is critical for establishing broadly applicable predictive models.
Second, our radiomic feature extraction relied on manual tumor segmentation, which may introduce observer variability. Although we performed inter-observer consistency analysis between two radiologists, automated segmentation algorithms could potentially provide higher reproducibility, particularly in defining tumor boundaries and peritumoral regions.
Third, while our model demonstrated favorable predictive performance, the ability to interpret the biological connections between these features and AKR1B10 expression remains limited. The absence of direct radio-pathological correlation analysis prevented us from confirming whether specific imaging features truly reflect AKR1B10-related histological alterations.
Fourth, our study focused exclusively on contrast-enhanced CT imaging without exploring other imaging modalities such as MRI, which may provide complementary information. Furthermore, we evaluated only a single molecular marker (AKR1B10) without considering broader combinations of molecular features, which might offer more comprehensive molecular subtyping of HCC.
Although our study confirmed a significant correlation between radiomic features and AKR1B10 expression, the underlying biological mechanisms remain to be elucidated. Future research should focus on exploring big data-based multiparametric MRI radiomics, and more wet laboratory studies are needed, including in vitro cell culture experiments and in vivo animal models (with controlled AKR1B10 expression levels), to validate whether AKR1B10-mediated metabolic alterations directly result in the observed imaging phenotypes.

7
Conclusion
We established AKR1B10 as a key prognostic marker and developed a novel CT radiomics model for its non-invasive prediction. The integrated nomogram provides a valuable tool for prognosticating HCC patients. Future work should focus on external validation in large-scale, multi-center cohorts, integration of multi-modal imaging data (e.g., MRI), and biological experimentation to elucidate the mechanistic pathways connecting AKR1B10 to its radiophenotype.

CRediT authorship contribution statement
Huang LiLi: Writing – review & editing, Formal analysis, Data curation. Hongan Ying: Formal analysis, Data curation, Conceptualization. Weiwen Hong: Writing – review & editing, Writing – original draft.

Consent for publication
Not applicable.

Ethics approval and consent to participate
Written informed consent was not required for this study because the data from TCGA and TCIA are deidentified and publicly available, obtaining written informed consent from individual patients is not necessary. Researchers must only agree to data use policies and cite the databases properly. This study received approval from the Ethics Committee of the First People's Hospital of Taizhou, Zhejiang Province, China, with the approval number 2023-KY047–01. All procedures adhered to the guidelines of the authors' institutional ethics committee and complied with the principles outlined in the Declaration of Helsinki.

Funding statement
This study was supported by the Fund of Taizhou Municipal Bureau of Science and Technology (Fund No.: 25ywb151 and 23ywa23).

Declaration of Competing Interest
The authors affirm no conflicts of interest and have provided their final approval, taking responsibility for all aspects of the work to ensure accuracy and integrity.