Automated Sinus MRI Phenotyping with PARASIDE: Construct-Specific Validation Against Independent Manual Annotations and Radiological Assessments in a Population-Based Cohort

Manual sinus magnetic resonance imaging (MRI) annotations are valuable but difficult to scale. We evaluated PARASIDE, an nnU-Net-based framework for automated sinus compartment segmentation on T1-weighted MRI, against historical manual annotations and radiological assessments generated independently of PARASIDE within the same longitudinal Study of Health in Pomerania (SHIP) imaging population. We linked 12,867 sinus sides from 3217 participants. Primary construct–marker pairings were specified before endpoint-specific modelling. Soft-tissue fraction discriminated abnormal sides in frontal (AUC 0.850, 95% CI 0.832–0.868) and maxillary sinuses (AUC 0.828, 95% CI 0.816–0.841). The topographic marker showed moderate discrimination of basal from apical opacification (AUC 0.746 left; 0.782 right), while total frontal volume discriminated aplasia/hypoplasia ratings (AUC 0.988); the near-absence threshold remained highly specific but insensitive. Primary findings were robust to exclusion of 279 participants overlapping with the original PARASIDE development population. Historical maxillary volumetric masks showed strong agreement, whereas frontal agreement improved after accounting for the predefined inferior extent of the non-equivalent historical segmentation approach. Polyposis remained weakly separable. PARASIDE showed correspondence with distinct, anatomically measurable MRI phenotypes across complementary historical reference layers, while morphology-specific constructs required dedicated reference standards.