CT-based CK19-oriented recursive nested habitat analysis for noninvasive phenotyping of hepatocellular carcinoma: a multicenter external validation study

Abstract Background Cytokeratin 19 (CK19)-positive hepatocellular carcinoma (HCC) represents an aggressive molecular subtype associated with invasive behavior and unfavorable outcomes. However, CK19 status is generally unavailable before treatment because it relies on pathological examination. This study aimed to develop and externally validate a CT-based imaging biomarker for noninvasive prediction of CK19 expression in HCC using recursive nested habitat analysis. Methods This retrospective multicenter study included 1029 patients with surgically treated HCC from three institutions. Preoperative contrast-enhanced CT images, tumor masks, clinicopathological characteristics, CK19 immunohistochemical status, and follow-up data were collected. A CK19-oriented recursive nested habitat framework was developed by modeling spatially localized tumor imaging patterns and progressively refining high-probability regions. Model performance was evaluated in internal and external validation cohorts. The incremental value of the habitat score beyond clinical variables was assessed, and exploratory analyses investigated associations with early recurrence and recurrence-free survival (RFS). Results Among 1029 patients, 459 (44.6%) were CK19-positive and 570 (55.4%) were CK19-negative. The recursive nested habitat score achieved robust discrimination of CK19 expression, with an external validation area under the receiver operating characteristic curve (AUC) of 0.808 (95% confidence interval [CI], 0.760–0.854). Adding the habitat score to clinical variables improved CK19 prediction compared with the clinical model alone. Prespecified ablation analyses demonstrated stable performance across different patch scales, enrichment fractions, and aggregation strategies. However, the nested habitat score did not provide significant incremental improvement for early recurrence prediction or maintain independent prognostic value for RFS after external validation. Conclusions A CT-based CK19-oriented recursive nested habitat framework provides a reproducible noninvasive imaging surrogate associated with CK19 expression in HCC across multicenter cohorts. The proposed model supports the potential of spatial CT phenotypes for molecular characterization of HCC, while recurrence analyses should be interpreted as secondary and exploratory. Further prospective studies integrating standardized imaging, pathology, and multi-omics data are required before clinical implementation.