Impact of algorithm choice when using off-the-shelf AI segmentation methods in large-scale medical analyses

Emelie Bäcklin, Eva Breznik

Keywords: image segmentation, computed tomography

TL;DR: Evaluation of the effect of lung lobe segmentation method on quantitative analysis of chest computed tomography

Abstract: Manual segmentation of medical images is time-consuming and impractical for large-scale studies. This study evaluated the impact of five deep learning–based lung lobe segmentation methods on lung volume and low attenuation volume below -950 Hounsfield units (LAV950) using chest CT scans from 2,579 participants in the SCAPIS cohort. Results showed minimal differences in mean lung volume and LAV950 across methods. AUC-ROC values for detecting chronic airflow limitation (CAL) were consistent for inspiratory CT but varied more for expiratory scans, indicating greater model sensitivity. Overall, segmentation choice had limited influence on downstream analysis of LAV950-derived lung function measures.

Serve As Reviewer: ~Mehdi_Astaraki1

Submission Number: 39