Conditional Latent Diffusion Models for PLAX Echocardiographic Image Synthesis: A Geometric-Anatomical Guided Approach

Abstract

Echocardiography is essential in cardiology, with the parasternal long axis (PLAX) view key to evaluating cardiac geometry and diagnosing ventricular hypertrophy. Measurements of cardiac structures—including ventricular dimensions and wall thickness, obtained by placing calipers on the walls—are often influenced by significant inter- and intra-clinician variability. The design of automated deep learning (DL) algorithms to address this issue and support clinician with measurements remains constrained by the limited availability of annotated echocardiographic datasets. To address this limitation, we propose a novel method to synthesize PLAX images by using spatial information from caliper positions. Our approach uses conditional latent diffusion models, guided by geometric-anatomical priors encoded through heatmap, ensuring alignment with cardiac anatomy. We evaluate the proposed method through a comprehensive comparison of conditioning strategies using specific metrics to quantify anatomical agreement between synthetic and real images. Our method shows superior performance, achieving the best image quality (Frechet Inception Distance:21.08) and anatomical consistency (Davies-Bouldin score:22.97). Synthetic images from our model enhanced the performance of DL models for caliper identification, showing their value in augmenting training datasets. By addressing variability and standardizing measurements, this approach paves the way for objective and reliable echocardiographic analysis, with promising implications for medical imaging analysis.