This retrospective exploratory study evaluated HCC patients with a binary Cs label in the study dataset. Ethical approval and the authors’ consent-waiver statement are provided in the Declarations (2025-E0322; 24 April 2025). We use “study-documented Cs status” to describe this project label, rather than a clinically adjudicated infection endpoint. Case-level diagnostic methods, assessment dates, treatment history, diagnostic certainty, and a negative reference standard were unavailable. The study used a binary study-documented Cs label. Although the historical project description cited WS309-2009, case-level diagnostic methods and the application of these criteria to individual patients could not be reconstructed.

Labels 1 and 0 are retained as dataset codes. Their operational contrast is documented Cs positivity vs. no documented Cs positivity; label 0 does not establish that infection was actively excluded, and label 1 does not establish current active infection. Counts by stool examination, serology, biliary examination, pathology or prior treatment, and a certainty-restricted analysis could not be obtained. These limitations also prevent determining the timing of Cs assessment relative to HCC diagnosis and imaging.

The analysis included 161 feature-complete patients: 77 with CT and 84 with MR imaging, with 103 label-1 and 58 label-0 patients. A fixed patient-level 80/20 split provided 128 training patients (61 CT, 67 MR) and 33 internal holdout patients (16 CT, 17 MR), with no overlap. Age, sex, hepatitis B virus (HBV)/hepatitis C virus (HCV) status, other liver-disease etiologies, cirrhosis, liver function, alpha-fetoprotein, tumor size, number and stage, histological differentiation, prior treatment, and imaging institution were unavailable in the analysis dataset. Clinical baseline comparisons and adjusted or combined clinical-radiomic models were therefore not estimable (see the clinical-variable availability Table in the Supplementary Materials).

The 33-patient holdout was reused across exploratory model comparisons. The responsible author confirmed that holdout results were not viewed before model selection; their possible influence on later manuscript decisions was not separately documented. We therefore treat holdout performance as internal exploratory evaluation. Additional MR sensitivity analyses used a different feature representation and split, described in Section “Sensitivity and robustness analyses”, and do not constitute independent validation.

CT/MR images and tumor masks were obtained from the existing study dataset. Each original pooled-model record represents one patient using their available imaging modality; the exact historical sequence-selection and aggregation rule could not be reconstructed. The original segmentation description cited nnU-Net with physician review^[28], and the available checkpoint was compatible. This discrepancy prevents definitive identification of the executed segmentation workflow. Reader expertise, correction frequency, blinding to Cs status, multifocal-lesion selection, and independent reference masks were undocumented. Accordingly, the available masks support feature extraction but do not establish segmentation accuracy or inter-reader reproducibility.

Cohort-wide CT phase, MR sequence/contrast phase, scanner and acquisition parameters were incomplete. Detailed Digital Imaging and Communications in Medicine (DICOM) metadata were available for one MR examination and cannot establish the settings used across the cohort. The exact historical extraction configuration, runtime and mapping between feature columns and ROI definitions were also incomplete. Available acquisition, preprocessing, and segmentation records, the original feature inventory, and the current MR extraction configuration are provided in Supplementary Methods. Current settings are distinguished from those used for the original pooled model.

Whole-tumor radiomics summarizes intensity and texture within the supplied tumor ROI. Habitat radiomics partitions the ROI into imaging-defined subregions using voxel-level features. CT and MR habitat models were fitted separately with training voxels, then applied to training and holdout images. In the patient-balanced construction, no more than 3,000 tumor voxels per training patient contributed to clustering, reducing the influence of large tumors. Actual training samples contained 145,174 CT and 110,876 MR voxels. Holdout voxels were excluded from centroid fitting.

Because habitat identities are modality-specific, pooled modeling used label-invariant summaries of per-habitat features, including central tendency, dispersion, extrema, and fraction-weighted quantities. These were concatenated with whole-tumor features. The saved fused representation contained 20 whole-tumor columns (19 radiomic descriptors and one voxel count) and 153 habitat-summary columns, totaling 173 columns for each evaluated cluster number (K). Each unsummarized habitat has the same 19 named radiomic descriptors plus its voxel count and tumor fraction (21 columns); the additional fraction, rather than an additional radiomic descriptor, explains the 20- vs.-21 difference. The full lists are supplied in Supplementary Methods. Habitat indices do not imply corresponding tissue compartments across CT and MR. The study design and image-derived workflow are shown in Figure 1.

Candidate pipelines combined training-side preprocessing, correlation pruning, feature selection and classification. Selectors included K-best, L1 regularization and elastic-net methods; classifiers included logistic regression, support vector machine (SVM), random forest (RF) and ExtraTrees. The training selection score was mean cross-validated area under the receiver operating characteristic curve (AUROC) minus 0.5 times its standard deviation. The patient-balanced combined model used K = 2, ExtraTrees, and elastic-net selection (enet10), with five-fold cross-validation (CV) repeated three times during training. Modality and acquisition identifiers were excluded from its predictor set.

K = 2 was retained from the original training-based model selection. It was among K = 2-6 evaluated for patient-balanced clustering; validity indices did not uniformly favor it, and the 173-column output dimension was constant across K. Thus, K = 2 reflects a model-development choice, without evidence of biological optimality or resampling stability. Whole-tumor, habitat-only, and combined alternatives are presented as descriptive comparisons on the reused holdout. Cluster-validity results for both the original K = 2-9 and patient-balanced K = 2-6 constructions are distinguished in Supplementary Figure 1 and Supplementary Table 1. The feature schemas are illustrated in Supplementary Figure 2. Model-search details are provided in Supplementary Methods; the full inventory is cited with the Results comparisons.

A modality-only control used CT/MR membership to assess discrimination attributable to modality composition on the same 33-patient holdout. An MR acquisition-only control used filename-derived protocol categories on the 17-patient MR holdout. These categories are acquisition proxies rather than verified contrast phases or complete scanner metadata. Paired image-minus-control comparisons tested whether the image model exceeded these limited acquisition descriptors.

CT and MR subgroup estimates were calculated from the same pooled image model. Separately trained modality-specific models were identified separately in the Supplementary Materials. Neither control adjusts for clinical covariates. AUROC was not estimated within an acquisition category containing only one study label.

Additional MR analyses used 107 original radiomic features from 207 matched image/mask pairs representing 106 patients. Features from multiple sequences were averaged within patient, without using labels or acquisition metadata as predictors. The MR split provided 90 development patients (70 label-1, 20 label-0) and 16 reserved patients excluded from these analyses. Of the 90 development patients, 69 also belonged to the original cohort, including 14 in its holdout; their labels agreed. This analysis uses a separate MR feature representation and does not reevaluate the original fused CT/MR habitat model.

Three repeats of five-fold outer CV used inner CV to select among four classifiers and five feature-count settings. Median imputation, standardization, and feature selection were fitted within resampling. Discrimination was summarized for each repeat and from each patient’s mean out-of-fold probability across repeats. Acquisition-signature controls and a radiomics analysis restricted to signatures containing both labels assessed sensitivity to acquisition composition. Habitat construction and K selection were not repeated. Batch harmonization was not performed because a defensible cohort-wide scanner or institution variable was unavailable.

Complementary analyses assessed contour perturbation, selection variability, and learning curves. The contour analysis used baseline masks and 16 synthetic variants on the 207 pairs. Forty patients contributed multiple pairs, so the pair-level intraclass correlation coefficient (ICC) intervals do not account for patient clustering and are not interpreted as patient-level inference. Feature sensitivity does not measure reader agreement or prediction robustness. Full extraction settings, perturbations, resampling procedures, and training sizes are provided in Supplementary Methods; detailed results are provided in the Supplementary Figures and Supplementary Tables.

Discrimination was summarized by AUROC with 95% confidence intervals (CIs) from 2,000 patient-level resamples, excluding one-class samples. The two acquisition-control comparisons used paired resampling. Two-sided bootstrap P values were twice the smaller zero-tail proportion with a one-count correction, capped at 1; seeds and numerical conventions are specified in Supplementary Methods. Estimates are conditional on the saved predictions and do not account for the complete model search or multiplicity. The comparisons are exploratory and use no confirmatory significance threshold. No prospective registration or sample-size calculation was documented.

Calibration and decision-curve analysis (DCA) were descriptive. The Brier score, expected calibration error (ECE), and Hosmer-Lemeshow statistic were reported without claiming validated clinical thresholds or utility. Precision was interpreted from the actual denominators and intervals rather than post-hoc observed-effect power. Feature-extraction and reporting context followed the radiomics literature^[29-31]. The additional MR analyses used Python 3.10.19, scikit-learn 1.7.2, and PyRadiomics 3.0.1; the complete available software versions and extraction settings are supplied in Supplementary Methods.

The 161-patient cohort included 103 label-1 and 58 label-0 patients. The internal holdout contained 21 label-1 and 12 label-0 patients: CT had 8/8 and MR had 13/4 label-1/label-0 patients. Among MR acquisition categories, MR_OTHER contained only label-1 patients (n = 10), whereas T1_VIBE_FS contained both labels (n = 7). Table 1 and Figure 2 show the cohort composition and the uneven acquisition support; Table 2 summarizes the holdout protocol support.

Holdout AUROCs were 0.690 for whole-tumor, 0.738 for habitat-only, 0.710 for the combined representation without patient-balanced reclustering, and 0.833 for the patient-balanced combined model. The latter had a training CV AUROC of 0.6915 ± 0.0887. Table 3 and Figure 3 show these descriptive comparisons; the shared holdout and model-search history prevent interpreting their ordering as independently established superiority.

Cluster-validity indices did not uniformly support K = 2 across modalities [Supplementary Figure 1 and Supplementary Table 1]. The original 173-column fused representation and the separate 107-feature MR representation are illustrated in Supplementary Figure 2.

The patient-balanced combined model achieved holdout AUROC 0.833 (95%CI: 0.683-0.955; n = 33). Subgroup AUROCs were 0.766 for CT (n = 16; 95%CI: 0.483-1.000) and 0.846 for MR (n = 17; 95%CI: 0.617-1.000). Figure 4 shows the empirical receiver operating characteristic (ROC) curve and subgroup AUROC estimates with intervals. The unweighted macro CT/MR AUROC was 0.806 [Supplementary Table 2]. The full model-search inventory is provided in Supplementary Table 3.

The modality-only control achieved AUROC 0.643; the paired image-minus-control difference was +0.190 (95%CI: 0.048-0.339; P = 0.008). Within MR, the acquisition-only control achieved AUROC 0.885, compared with 0.846 for the pooled image model; the paired difference was -0.038 (95%CI: -0.232-0.115; P = 0.742). Thus, discrimination exceeded simple modality membership but did not exceed the MR acquisition proxy. Figure 5 and Table 4 present both comparisons.

The pooled image model and modality-only control were compared in the same 33-patient historical internal holdout. Within the 17-patient MR subset, predictions from the same pooled image model were compared with a filename-derived MR acquisition-proxy control. The paired image-minus-control differences in area under the receiver operating characteristic curve (ΔAUROC) were +0.190 (95%CI: 0.048-0.339; P = 0.008) for the modality comparison and -0.038 (95%CI: -0.232-0.115; P = 0.742) for the MR acquisition-proxy comparison. These analyses are exploratory and are based on the reused historical holdout. The acquisition categories are filename-derived internal proxies and do not establish verified imaging protocols or contrast phases. Labels 0 and 1 denote study-documented Cs status; label 0 does not establish active exclusion of infection, and label 1 does not establish active infection.

AUROC was not estimable within the one-class MR_OTHER category (10 label-1 and 0 label-0 patients). Exploratory T1_VIBE_FS sensitivity analyses are reported separately in Supplementary Table 4. Descriptive calibration and DCA results are provided in Supplementary Figures 3 and 4, and Supplementary Table 5; given the small reused holdout, these analyses are not interpreted as evidence of validated calibration, clinical threshold selection, or clinical utility.

In the separate 90-patient MR development analysis, AUROC from probabilities averaged across three nested CV repeats was 0.8536 (95%CI: 0.7650-0.9257), with repeat-specific AUROCs from 0.7550 to 0.8771. Acquisition-only controls achieved AUROCs of 0.8000, 0.8004, and 0.7829 for ExtraTrees, logistic regression, and RF, respectively. Only 34 patients (18 label-1, 16 label-0) had acquisition signatures containing both labels. Radiomics AUROC in this subset was 0.6840 (95%CI: 0.4861-0.8523). Figure 6 summarizes the resampling and acquisition findings.

Across 24 features linked to saved MR models, median contour-perturbation ICCs ranged from 0.8613 to 0.9713 and median relative changes from 0.0141 to 0.1126. Median pairwise feature-selection Jaccard was 0.2991 [interquartile range (IQR): 0.2991-1.0000], indicating variation in selected features. Learning-curve performance was nonmonotonic across training sizes of 18-72 patients. Feature-family details and unavailable clinical variables are documented in Supplementary Tables 6 and 7. Detailed contour, feature-selection, learning-curve, and acquisition-support results are in Supplementary Figures 5-7 and Supplementary Table 8. Supplementary Figure 8 summarizes the available supporting records and their scope.

Whole-tumor and habitat radiomics showed an exploratory association with study-documented Cs status in this retrospective HCC cohort. Three findings frame its interpretation. First, the combined image model discriminated the study labels on internal evaluation, although the estimate was imprecise. Second, it exceeded a modality-only control but did not exceed the acquisition-only control within MR. Third, the separate MR analysis showed lower discrimination after restriction to acquisition signatures containing both labels. Together, these observations suggest that image-derived associations in this dataset cannot be separated confidently from acquisition composition and endpoint uncertainty.

The positive paired comparison with CT/MR membership indicates information beyond that single descriptor. It does not establish incremental clinical value, because liver-disease etiology, tumor burden, and other clinical predictors were unavailable. Likewise, the difference between whole-tumor and combined models cannot establish a benefit from habitats after extensive exploratory model development. The strongest conclusion is an observed association whose clinical and biological specificity remains unresolved.

The closest disease-specific reference is Liu et al., who studied 134 HCC patients with gadoxetic acid-enhanced MR imaging and developed radiomics models for liver-fluke infection and cytokeratin 7/19 expression^[10]. That work provides a relevant disease and imaging context, but differences in endpoint ascertainment, pooled CT/MR composition, and acquisition controls limit direct performance comparisons.

Bernatowicz et al. evaluated voxel-wise radiomics repeatability and reproducibility in lung CT, including test-retest data, and showed why habitat construction can be sensitive to feature stability^[23]. Fang et al. used multi-habitat MR imaging signatures to predict treatment response in locally advanced cervical cancer^[24]. Shang et al. evaluated tumoral and peritumoral habitat radiomics in a large multicenter lung CT study with internal validation and external test sets^[25]. These studies support the methodological relevance of spatial heterogeneity and reproducibility assessment. Their disease sites, endpoints, acquisition settings, and validation structures differ from the present exploratory association with study-documented Cs status, so no priority or superiority claim is made.

Habitat summaries offer a way to describe spatial variation that whole-tumor averages can obscure. Chronic hepatobiliary inflammation, epithelial injury, biliary remodeling, and immune-context differences are plausible links between Cs exposure and imaging phenotype^[4,6,8,9], but those mechanisms were not measured in this cohort. Separate CT and MR clustering and label-invariant summaries avoid equating habitat indices across modalities, but they do not identify necrosis, inflammation, or a Cs-specific tissue process without histologic correspondence. Contour sensitivity and variation in feature selection further caution against assigning biological meaning to a single fitted feature ranking.

Acquisition controls are especially informative when exposure labels are unevenly distributed across imaging categories. Here, the MR acquisition control discriminated the labels at least as well as the pooled image model, while several signatures contained only one class. Restriction to shared support reduced the sample and the observed AUROC. This pattern is compatible with acquisition-label coupling; it does not prove that acquisition alone caused the image association, because restriction also changes the patient sample and yields a wide interval.

Case-level Cs ascertainment, timing, treatment history, and a verified negative reference standard were unavailable. Study labels may mix active infection, previous infection, and classification error. Missing clinical covariates prevent adjustment for major HCC etiologies, liver disease and tumor burden. The results therefore cannot establish infection diagnosis or a Cs-specific imaging phenotype.

The internal holdout was small, reused across model comparisons; protection from all later analytical decisions could not be established. Its performance may be optimistic relative to independent external validation. Its AUROC exceeded the selected model’s training CV mean, and conditional bootstrap intervals do not capture the full model-selection uncertainty. The separate MR development cohort included 14 original holdout patients and used a different feature representation; its nested estimates neither independently validate nor correct the original habitat-model result. No external, temporal, prospective, or center-held-out cohort was available.

Acquisition metadata were incomplete, the MR holdout contained only four label-0 patients, and one acquisition category contained a single class. The shared-support estimate was also imprecise. Neither the controls nor subgroup analyses establish scanner-, sequence-, or institution-independent performance.

The historical segmentation description and available checkpoint could not be fully reconciled. Reader expertise, correction practices, blinding, and multifocal-lesion selection were undocumented. The exact original extraction configuration and feature-to-ROI mapping were incomplete. Synthetic perturbations assessed feature sensitivity rather than reader agreement or model prediction robustness; repeated image/mask pairs were not accounted for in pair-level uncertainty estimates. These limitations restrict reproducibility and biological interpretation.

Future evaluation should prioritize clinically adjudicated Cs endpoints, linked clinical covariates, complete acquisition and segmentation documentation, patient-level isolation and prespecified model development, followed by independent validation. Acquisition-balanced or sequence-matched evaluation would help determine whether the association persists beyond imaging practice. Calibration and decision curves from the current small reused holdout cannot establish clinical utility.

Whole-tumor and habitat radiomics were associated with study-documented Cs status in this retrospective HCC cohort. Discrimination exceeded a modality-only control but did not exceed an MR acquisition-only control, and the separate MR shared-support analysis yielded a lower AUROC. These findings support a narrow exploratory association under substantial acquisition coupling and endpoint uncertainty, requiring clinically verified labels, clinical adjustment and independent evaluation before diagnostic interpretation.