Establishment of normative spleen volume in a Japanese cohort using automated CT segmentation: diagnostic...

BACKGROUND: In this study, we aimed to establish specific reference values for spleen volume in Japan using automated deep learning-based segmentation as well as evaluate the diagnostic utility of volume-defined splenomegaly in staging liver fibrosis. METHODS: We analyzed 4,732 healthy Japanese adults (Data Set 1) who underwent unenhanced abdominal computed tomography. Spleen volumes were measured using a deep learning-based segmentation tool (mean Dice coefficient = 0.95; median processing time = 26.6 s). A multivariable linear regression model incorporating sex, age, weight, and height was developed and validated. An external cohort of 136 patients with histologically confirmed liver disease (Data Set 2) was used to assess the clinical utility of volume-based splenomegaly. Z-scores were calculated from prediction residuals, using thresholds of Z ≥ + 1 and Z ≥ + 2. Diagnostic performance for cirrhosis (F4) was evaluated using receiver operating characteristic analysis. RESULTS: The final predictive model used weight alone (Predicted spleen volume = 3.08 + 1.96 × Weight [kg]), with a mean absolute error of 38.5 mL. In Data Set 2, splenomegaly prevalence increased with fibrosis stage: for Z ≥ + 1, it was 0.0% (F1), 32.7% (F2), 54.3% (F3), and 74.4% (F4); and for Z ≥ + 2, it was 0.0%, 16.3%, 34.3%, and 48.8%, respectively. The area under the curve for detecting F4 was 0.73. CONCLUSION: We established population-specific spleen volume reference values using deep learning-based segmentation. The model enables objective identification of splenomegaly from routine clinical parameters, supporting its integration into artificial intelligence-assisted radiology workflows.