LDM-Echo: A Latent Diffusion Model for Echocardiographic Image Denoising with Clinical Evaluation

Speckle noise in echocardiographic imaging degrades contrast, obscures anatomical structures, and compromises quantitative measurements. Effective suppression of this noise is critical, yet traditional denoising methods often cause structural loss through over-smoothing. To address these challenges, LDM-Echo is proposed, a novel denoising framework based on a Latent Diffusion Model (LDM) that operates in a compressed latent space learned via a convolutional autoencoder. The autoencoder is trained with a hybrid loss function combining MSE, SSIM, and LPIPS to ensure high-fidelity and perceptually coherent reconstruction. Denoising is formulated as a controlled reverse diffusion process conditioned on the noisy latent representation, enabling effective noise suppression while preserving diagnostically relevant anatomical structures at substantially lower computational cost than pixel-space diffusion. The model was trained on the CAMUS dataset and evaluated across multiple degradation levels using conventional image quality metrics (PSNR, SSIM, LPIPS, and FID) as well as ultrasound-specific clinical metrics (CNR, ENL, and CoV). To assess clinical utility, we further performed downstream validation through automated cardiac segmentation and Left Ventricular Ejection Fraction (LVEF) estimation. Generalization was evaluated on the unseen EchoNet-Dynamic dataset without fine-tuning using reference-free robustness metrics (SSI, GCI, and Entropy). The proposed framework achieves a PSNR of 33.57 dB, an SSIM of 0.9127, and an LPIPS of 0.0219 at moderate noise levels, outperforming classical, CNN-based, GAN-based, and pixel-space diffusion approaches in perceptual quality. Patient-level clinical validation on poor-quality images yields ENL gains of up to 9.7% and CoV reductions of 4.5%. Furthermore, downstream segmentation and functional evaluation demonstrated improved clinical assessment, reducing Left Ventricular Ejection Fraction (LVEF) estimation error from 6.36% ± 7.12% to 5.82% ± 4.66% (MAE) and increasing correlation with expert measurements from 0.586 to 0.755. Cross-dataset evaluation on 3000 EchoNet images confirms strong speckle suppression (ENL +117.9%, SSI -32.3%, CoV -33.7%) while preserving overall image characteristics, demonstrating robust generalization without domain adaptation.