AI deep learning model quantified aortic valve calcium in echocardiography with AUROC 0.911 and R2 up to 0.721, maintaining accuracy across unseen centers.
Does a deep learning model applied to transthoracic echocardiography accurately quantify aortic valve calcium score compared to computed tomography?
A deep learning model can accurately quantify aortic valve calcium score using standard transthoracic echocardiography, demonstrating excellent performance and generalizability.
Absolute Event Rate: 0% vs 0%
Abstract Background Aortic valve calcification (AVC) plays a key role in the progression of aortic valve stenosis. Computed tomogaphy (CT) aortic valve calcium score (AVCS) is the gold standard for AVC quantification. Purpose To develop deep learning models for the quantification of AVC in echocardiography (TTE). Methods A total of 471 TTE studies and the corresponding AVCS (Agatston score, Agatston unit AU) obtained from 365 CT from 234 patients and 9 centres were available. The dataset was split at a patient level into 194 TTE studies (1189 videos) for training and validation and 274 studies (1484 videos) for testing, maintaining 89 studies (434 videos) from 2 centres solely in the testing set to evaluate generalizability. The models were trained by fine-tuning the PanEcho 1 network with added specific heads for two main tasks: differentiation between AV with and without AVC and the numerical AVCS quantification, and two ancillary tasks, i.e. two multiclass classifications (0, 1-895 and 895 AU; and 0, 1-2951 and 2951 AU) 2. AVCS prediction were set to zero when all classifications predicted the absence of AVC. Five-fold cross-validation was used for training, generating five models for each task from parasternal long and short axis view and 3-chamber view. After outliers removal, the mean predictions of the 5 models was obtained and compared with AVCS. Results Patients had a median inter-quartile age of 48.3 39.2; 63.3, 63 (27%) were female and 180 (77%) had a bicuspid aortic valve. The cohort included 108 (46%) patients without AVC while those with calcium presented a AVCS of 818.6 289.7; 2025.5. The time difference between TTE and CT was 35.2 4.1; 124.5 days. The model differentiating patients with and without AVC had excellent accuracy in internal cross-validation (AUROC 0.994, 95% confidence interval 0.984-1.000) (figure 1). In the testing set, which included data from unseen centres, the model had excellent discriminatory performance in differentiating patients with and without AVC (AUROC = 0.911 0.875-0.942) (figure 2). Of note, the performance was maintained in TTE from centres excluded from the training set (AUROC = 0.877 0.802-0.939), highlighting its generalizability. Model for AVCS quantification were excellent in cross-validation (R2 = 0.820, median and IQR of error 954 211.6; 2368.6 AU), testing set (R2 = 0.721, error 229 0;967 AU) and TTE studies from unseen clinical centres (R2 = 0.637, error 98 0;612 AU). No difference was seen comparing errors in TTE from centres included or not in the training (p=0.107, Mann-Whitney U test) nor between vendors (316 studies from General Electric and 155 from Philips, p=0.288) or sexes (p=0.645). Conclusions A deep learning model effectively quantified aortic valve Agatston score in transthoracic echocardiography, demonstrating excellent performance and generalizability. These findings support the potential role of AI-assisted echocardiography for aortic valve calcium assessment.Figure 1.Cross-validation performance Figure 2 External validation performance
López-Gutiérrez et al. (Sat,) reported a other. AI deep learning model quantified aortic valve calcium in echocardiography with AUROC 0.911 and R2 up to 0.721, maintaining accuracy across unseen centers.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: