medRxiv [Preprint]. 2026 Sep 15:2026.09.14.26363024. doi: 10.64898/2026.09.14.26363024.
ABSTRACT
The limited availability of imaging data for uncommon cardiovascular phenotypes constrains the development of robust imaging models. We evaluated whether class-conditional diffusion models can generate synthetic transthoracic echocardiograms that improve downstream cardiac imaging tasks. The primary application was cardiac amyloidosis detection in a Duke University cohort using a two-step classifier, with external analyses using EchoNet-Dynamic for image-fidelity assessment and EchoNet-LVH for wall-thickness phenotype classification. The Duke cohort was partitioned at the patient-encounter level into 70% training, 15% validation, and 15% test sets; generators were trained only on the training partition, augmentation levels were selected using validation AUROC, and final evaluation used held-out real test data. In the Duke all-view two-step analysis, adding synthetic images increased AUROC from 0.883 to 0.924, with an AUROC difference of 0.041 (95% CI, 0.013-0.069); in EchoNet-LVH, AUROC increased from 0.832 to 0.864, with an AUROC difference of 0.032 (95% CI, 0.021-0.044). Expert review found that synthetic images were sometimes difficult to identify as synthetic, but rated them lower for diagnostic adequacy. These findings suggest that diffusion-generated echocardiograms may provide a practical approach to augmenting limited training data for selected cardiac imaging tasks and motivate further evaluation across clinical settings.
PMID:42779795 | PMC:PMC13596512 | DOI:10.64898/2026.09.14.26363024