Inter-observer variability of expert-derived morphologic risk predictors in aortic dissection

Objectives Establishing the reproducibility of expert-derived measurements on CTA exams of aortic dissection is clinically important and paramount for ground-truth determination for machine learning. Methods Four independent observers retrospectively evaluated CTA exams of 72 patients with uncomplicated Stanford type B aortic dissection and assessed the reproducibility of a recently proposed combination of four morphologic risk predictors (maximum aortic diameter, false lumen circumferential angle, false lumen outflow, and intercostal arteries). For the first inter-observer variability assessment, 47 CTA scans from one aortic center were evaluated by expert-observer 1 in an unconstrained clinical assessment without a standardized workflow and compared to a composite of three expert-observers (observers 2-4) using a standardized workflow. A second inter-observer variability assessment on 30 out of the 47 CTA scans compared observers 3 and 4 with a constrained, standardized workflow. A third inter-observer variability assessment was done after specialized training and tested between observers 3 and 4 in an external population of 25 CTA scans. Inter-observer agreement was assessed with intraclass correlation coefficients (ICCs) and Bland-Altman plots. Results Pre-training ICCs of the four morphologic features ranged from 0.04 (-0.05 to 0.13) to 0.68 (0.49-0.81) between observer 1 and observers 2-4 and from 0.50 (0.32-0.69) to 0.89 (0.78-0.95) between observers 3 and 4. ICCs improved after training ranging from 0.69 (0.52-0.87) to 0.97 (0.94-0.99), and Bland-Altman analysis showed decreased bias and limits of agreement. Conclusions Manual morphologic feature measurements on CTA images can be optimized resulting in improved inter-observer reliability. This is essential for robust ground-truth determination for machine learning models.

Automated summary

This study aimed to evaluate the reproducibility of expert-derived CTA measurements. The results showed that by optimizing the manual measurement method, inter-observer consistency can be improved, which is crucial for the ground-truth determination of machine learning models.