Shaping Quality in Cardiovascular Magnetic Resonance: A Comparative Study of Segmentation Approaches by Trainees and Experts

Abstract

Background

Cardiovascular magnetic resonance (CMR) is an established cardiovascular imaging (CVI) technique. Deficits in training limit the widespread use of CMR. This study analyzed the influence of CVI experience on segmentation, to define quality standards for teaching and supervision.

Methods

Four CMR experts determined left ventricular (LV) and right ventricular (RV) gold-standard contours in end-systole (ES) and end-diastole (ED), by consensus. After a brief teaching session, readers independently performed segmentations. Readers were classified as beginners (no previous experience in CVI), intermediates (previous experience in CVI, but not in CMR), or experts (extensive experience in CVI including CMR). Results were compared, and the cause of deviation was analyzed, using metrics such as the Dice similarity coefficient (DSC).

Results

A total of 46 readers (19 beginners, 21 intermediates, 6 experts) performed image analysis. Using the DSC, we found significant differences in endocardial LV ED contours (median [interquartile range]: beginners, 92.9% [91.9%-93.5%]); intermediates, 93.5% (93.0%-94.1%); experts, 93.9% (93.1%-94.3%); P = 0.043) and in myocardial contours (beginners, 79.0% (75.0%-80.9%); intermediates, 80.9% (78.0%-82.4%); experts, 85.0% (79.8%-86.5%); p = 0.001). Experts had higher DSC scores for the right ventricle (ES: beginners, 83.8% (81.3%-85.8%); intermediates, 81.7% (79.6%-85.6%); experts, 89.0% (86.6%-89.8%); P = 0.003; ED: beginners, 89.2% (88.1%-90.3%); intermediates, 88.6% (87.9%-89.2%); experts, 91.6% (89.8%-93.3%); P = 0.002). The disagreements were not traceable in absolute volume and function (P for all > 0.2). Sources of disagreement were related mainly to handling of basal slices.

Conclusions

After a brief standardized teaching session, beginners and intermediates performed chamber quantification consistent with that of experts. Differences, especially in LV mass and RV segmentations, warrant continuous training, ideally accompanied by automatic methods for quality assurance.