CT Ventricular Volume Reference Percentile Curves Using Automated nnU-Net Segmentation: A Study with LLM-Facilitated Cohort Selection

Background: Age- and sex-specific CT reference percentile curves of ventricular volume may support objective assessment of ventricular enlargement. We aimed to generate such curves by combining a locally deployed large language model (LLM) for cohort curation with nnU-Net 12-subcompartment segmentation. Methods: This retrospective study did not require formal ethics approval under the Swiss Human Research Act. We screened 42,211 head CT reports from January 2019 to April 2024. A phrase prefilter identified 9622 candidates for German-tuned 70-billion-parameter LLM review of stroke, leukoencephalopathy, parenchymal defects, enlarged ventricles, and space-occupying lesions; 6516 remained. A second LLM workflow screened 210,761 clinical data warehouse summaries to exclude dementia, organic brain disease, or cognitive decline. After image-availability and series-mismatch exclusions, the final cohort comprised 2608 CTs from 2608 patients (1374 males, 1234 females; age 14–98 years). Ventricular masks were obtained using a validated nnU-Net tool integrated in TotalSegmentator. Results: LLM labelling accuracy in 471 blinded reports ranged from 0.966 to 0.992. CDW screening audit showed 91% agreement, sensitivity 0.96, and negative predictive value 0.96. Median total ventricular volume rose from 10.94 mL in young females and 13.42 mL in young males to 21.97 mL and 31.97 mL, respectively, in the 60–69 age group. Except for the fourth ventricle, compartments showed positive age correlations (r = 0.44–0.69); the fourth ventricle showed no relevant association (r = 0.01 males; r = 0.05 females). After adjustment for age, male sex remained associated with greater total ventricular volume (6.54 mL; 95% CI, 5.61–7.46; $p<0.001$). Conclusions: This study provides age- and sex-specific CT ventricular nomograms in a selected reference cohort. The LLM-assisted curation and nnU-Net segmentation workflow provides a feasible approach for large-scale cohort analysis.