AI Reads Liver Scans to Tell Hepatitis B Cirrhosis Apart from Schistosomiasis

In regions where both hepatitis B virus and schistosomiasis are endemic, two patients can arrive at a clinic with livers that look equally scarred and hardened, yet carry diseases that demand entirely different treatments. Cirrhosis driven by the hepatitis B virus arises from chronic viral infection and immune-mediated destruction of liver tissue, while cirrhosis caused by schistosomiasis results from parasitic eggs lodging in the portal venous system and triggering fibrosis around the blood vessels. Distinguishing between the two etiologies has long required invasive liver biopsy or reliance on clinical and laboratory clues that are not always conclusive. Now a multicenter research team in China has shown that a machine-learning analysis of routine computed tomography scans, combined with basic clinical data, can make that distinction non-invasively with remarkable accuracy.

The study, published in BMC Medical Imaging, enrolled 143 patients with either hepatitis B virus-related cirrhosis or schistosomiasis-related cirrhosis across three institutions between August 2019 and December 2021. Rather than asking radiologists to eyeball the scans, the researchers applied a technique called radiomics, which converts medical images into hundreds of quantitative features that capture subtle patterns of texture, shape, and intensity invisible to the human eye. These features were extracted from non-contrast CT images of the liver, meaning patients did not need contrast agents or additional scanning procedures beyond the imaging they would receive anyway.

To build a reliable signature from this high-dimensional feature space, the team used a two-stage statistical pipeline. First, minimum redundancy maximum relevance, or mRMR, screened the extracted features to identify those most informative about the disease etiology while avoiding redundant overlap. Then, least absolute shrinkage and selection operator regression, known as LASSO, further compressed the feature set and assigned each surviving feature a weighted coefficient, producing a single composite score called the Rad-score. This score summarizes the aggregate imaging fingerprint of each patient’s liver into one number that reflects the probability of one etiology versus the other.

The results were striking. The Rad-score was significantly lower in patients with hepatitis B virus-related cirrhosis than in those with schistosomiasis-related cirrhosis, with P values below 0.001 in both the training and internal validation cohorts. This suggests that the two diseases leave measurably different traces in the texture and structure of liver tissue on plain CT scans, traces that the algorithm can detect even when the overall appearance of cirrhosis looks similar. Schistosomiasis-related fibrosis, which concentrates around portal vein branches and produces a distinctive periportal pattern, appears to generate a different radiomic texture than the diffuse parenchymal damage typical of chronic viral hepatitis.

Building on the imaging signature, the researchers developed three diagnostic models using multivariable logistic regression: a clinical model based on patient characteristics alone, a radiomics model based on the Rad-score alone, and an integrated clinical-radiomics model combining both sources of information. The integrated model proved the strongest performer, achieving an area under the receiver operating characteristic curve, or AUC, of 0.990 in the training cohort, 0.920 in the internal validation cohort, and 0.782 in the external validation cohort. An AUC of 0.990 means the model could correctly rank a randomly chosen pair of patients, one from each disease group, nearly every time in the training data.

Performance naturally declined when the model was tested on data from institutions it had never seen, dropping to 0.782 in external validation, a pattern common in radiomics research where scanner settings, patient demographics, and imaging protocols differ across sites. Notably, however, both the radiomics and integrated models showed more consistent performance across the internal and external validation cohorts than the clinical model alone, suggesting that the imaging features carry transferable information about disease etiology that clinical variables by themselves do not capture. The integrated model also demonstrated the greatest overall clinical utility in decision curve analysis, a method that evaluates whether a model would benefit patients across a realistic range of diagnostic thresholds rather than at a single arbitrary cutoff.

Calibration curves further confirmed that the integrated model’s predicted probabilities tracked well with observed disease frequencies, an important property for clinical deployment because a model that is accurate on average but poorly calibrated can mislead individual treatment decisions. Decision curve analysis complemented this by showing that, across a broad span of threshold probabilities, acting on the integrated model’s predictions would yield more net benefit than treating all patients as having one etiology or the other, or than relying on the clinical model in isolation.

The clinical implications are considerable for endemic regions where schistosomiasis and hepatitis B coexist. Accurate etiological differentiation matters because the two conditions differ substantially in pathogenesis, clinical management, and prognosis. Schistosomiasis-related cirrhosis may respond to antiparasitic therapy and management of portal hypertension, while hepatitis B-related cirrhosis requires antiviral treatment and surveillance for hepatocellular carcinoma. A non-invasive tool that can be applied to CT scans already obtained for routine care, interpreted alongside standard clinical and laboratory evaluation, could spare patients unnecessary biopsies and accelerate appropriate treatment pathways.

The authors emphasize that their model is intended as a complementary imaging tool rather than a replacement for clinical judgment, and that further prospective multicenter studies are warranted to validate its utility before widespread adoption. The retrospective design, the relatively modest sample size of 143 patients, and the performance drop in external validation all point to the need for larger, prospectively designed trials. Nevertheless, the study adds to a growing body of evidence that quantitative image analysis can extract diagnostic information from routine scans that far exceeds what conventional radiological interpretation provides, potentially transforming how etiological assessment of cirrhosis is conducted in resource-limited endemic settings.

Subject of Research: CT-based radiomics combined with clinical features for non-invasive differentiation of hepatitis B virus-related and schistosomiasis-related cirrhosis

Article Title: CT-based radiomics combined with clinical features for non-invasive differentiation of hepatitis B and schistosomiasis cirrhosis: a multicenter study

Article References: Tong, J., Xu, X., Hu, F., Wang, Z., Ma, L., Li, J., Li, X., Yang, Z., & Xia, X. (2026). CT-based radiomics combined with clinical features for non-invasive differentiation of hepatitis B and schistosomiasis cirrhosis: a multicenter study. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02693-6

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02693-6

Keywords: radiomics, hepatitis B virus, schistosomiasis, liver cirrhosis, computed tomography, clinical-radiomics model, non-invasive diagnosis, machine learning, BMC Medical Imaging, multicenter study, CT-based, combined