Investigation of Concordance between Artificial Intelligence and Manual Measurements of the Cardiac Parameters in Cardiac Magnetic Resonance Imaging.
Authors
Affiliations (4)
Affiliations (4)
- Department of Radiology, The Central Hospital of Wuhan, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, 430014, China.
- Department of Radiology, West China Hospital, Sichuan University, Chengdu, 610041, China.
- Department of Cardiology, West China Hospital, Sichuan University, Chengdu, 610041, China. [email protected].
- Department of Radiology, The Central Hospital of Wuhan, Tongji Medical College, Huazhong University of Science and Technology, Wuhan, 430014, China. [email protected].
Abstract
Commercial artificial intelligence (AI) software for cardiac magnetic resonance (CMR) analysis has shown promising internal validation results, but independent external validation in distinct clinical settings remains limited. This study aimed to externally validate a commercially available AI system for biventricular parameters and per-segmental left ventricular wall thickness (LVWT) measurements in a prospective cohort independent from the training data. In this prospective study, 100 healthy Chinese volunteers (67 men; age 20-63 years) from a single center (The Central Hospital of Wuhan) underwent 3.0 T CMR with cine short-axis and long-axis sequences. Biventricular parameters and per-segmental LVWT were measured by commercial AI software (SHUKUN) and a single senior cardiac radiologist with 9 years of experience. Paired t tests, Wilcoxon rank-sum tests, interclass correlation coefficient (ICC), and Spearman correlation were performed. Bland-Altman and Bull's-eye plots were generated. AI and manual measurements showed good to excellent consistency and correlation for most biventricular parameters in males and the entire cohort (ICCs = 0.551-0.922; r = 0.656-0.910; all p < 0.05), except right ventricular ejection fraction (r = 0.374, ICC = 0.330). Female subjects showed greater variability. For per-segmental LVWT, most segments demonstrated good agreement (ICCs = 0.415-0.776; r = 0.436-0.794; all p < 0.05), with suboptimal performance in anterior and apical segments. This independent external validation demonstrates that commercial AI software has good consistency with manual measurements for most cardiac parameters in healthy adults. However, manual verification remains necessary for segmental LVWT analysis, particularly in female subjects. Validation in diseased populations is warranted before clinical implementation.