In Journal of cardiovascular computed tomography
BACKGROUND : Quantitative coronary plaque parameters are increasingly being utilized as surrogate endpoints of pharmaceutical trials. However, little is known whether differences in segmentation significantly alter parameter values.
METHODS : Overall, 100 coronary plaques with adverse imaging characteristics were segmented automatically, by two experts (R1-R2) and three nonexperts (R3-R5). Low attenuation noncalcified (LANCP), noncalcified and calcified plaque volume were calculated and 4310 radiomic features were extracted. Intraclass correlation coefficient (ICC) values were calculated between the segmentations.
RESULTS : ICC values between expert readers were 0.84 [CI: 0.77-0.89] for total; 0.83 [CI: 0.76-0.88] for noncalcified; 0.96 [CI: 0.94-0.98] for calcified and 0.65 [CI: 0.51-0.75] for LANCP volumes. Comparing nonexperts' and experts' results, ICC ranged between 0.64 and 0.90 for total; 0.63-0.91 for noncalcified; 0.86-0.96 for calcified and 0.34-0.84 for LANCP volume. All readers (R1-R5) showed poor agreement with automatic segmentation (range: 0.00-0.27), except for calcified plaque volumes (range: 0.73-0.88). Regarding radiomic features, expert readers (R1-R2) achieved good reproducibility (ICC>0.80) in 88.6% (39/44) of first-order, 62.0% (424/684) of gray level co-occurrence matrix (GLCM), 75.8% (50/66) of gray level run length matrix (GLRLM) and 19.8% (696/3516) of geometrical parameters. Between experts and nonexperts, ICC ranged between: 70.5%-86.4% for first-order, 31.0%-58.3% for GLCM, 24.2%-78.8% for GLRLM and 6.2%-21.1% for geometrical features, while between all readers and automatic segmentation ICC ranged between: 25.0%-38.6%; 0.0%-0.0%; 0.0%-3.0% and 1.1%-1.4%, respectively.
CONCLUSIONS : Even among experts there is a considerable amount of disagreement in LANCP volumes. Nevertheless, expert readers have the best agreement which currently cannot be replaced with nonexperts' or automatic segmentation.
Kolossváry Márton, Jávorszky Natasa, Karády Júlia, Vecsey-Nagy Milán, Dávid Tamás Zoltán, Simon Judit, Szilveszter Bálint, Merkely Béla, Maurovich-Horvat Pál
Artificial intelligence, Coronary plaque, Machine learning, Radiomics, Segmentation