Image registration improves inter-reader agreement of objective response in CT assessment of pancreas adenocarcinoma

Objectives Pancreatic ductal adenocarcinoma (PDAC) exhibits ill-defined appearance on imaging during neoadjuvant therapy (NAT), which may limit consistency of dimensional measurements for objective response. This study evaluates inter-reader variability of current clinical measures of tumor size and interval response between baseline and restaging CT, and compares reader reliability with and without image registration assistance. Methods Two senior and two junior radiologists retrospectively reviewed imaging of resectable PDAC patients enrolled in an IRB-approved Phase II clinical trial. Readers segmented lesion volumes from baseline and restaging CT and measured axial plane diameters to evaluate therapeutic response according to RECIST v1.1 and WHO criteria. Reliability of criteria index measures were compared with a computer-assisted image registration approach. Inter-reader agreements were quantified via generalized conformity index (GCI), concordance correlation coefficient (CCC), and Fleiss’ Kappa (k). Statistical analyses were performed with significance level α = 0.05. Results Baseline and restaging imaging of 30 patients (67 ± 10y, 50% female) were evaluated. Inter-reader agreements of segmented tumor volumes at baseline/restaging were low (junior: GCI = 0.42/0.35, CCC = 0.36/0.68; senior: GCI = 0.43/0.39, CCC = 0.62/0.96; all readers: GCI = 0.43/0.38, CCC = 0.66/0.78). Agreements in percent change in tumor size differed by experience level (longest diameter: CCC = 0.26/0.77/0.59, p ' 0.01; diameter product: CCC = 0.32/0.86/0.66, p ' 0.001) and produced low kappa agreements (RECIST: k = 0.0/0.37/0.23, p ' 0.01; WHO: k = 0.0/0.31/0.18, p = 0.01 [junior/senior/all readers]). Image registration-assisted tumor response produced higher agreements across all levels of reader experience (k = 0.82/0.80/0.81, p ' 0.01; CCC = 0.83/0.95/0.91, p ' 0.05 [junior/senior/all readers]). Conclusion Substantial measurement variability exists in CT assessment of PDAC tumor response during NAT. Longitudinal image registration improves reproducibility and diminishes experience gap in radiologist assessments of objective response.