UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Synthetic echocardiogram generation with conditional diffusion models for segmentation tasks Ashrafian, Pooria

Abstract

The availability of high-quality, annotated training data critically influences the effectiveness of deep neural networks in echocardiographic image analysis. However, acquiring such data at scale remains challenging due to the need for expert clinical annotation and the variability introduced by different ultrasound vendors, devices, and acquisition protocols. These factors hinder the robustness and generalizability of segmentation models across diverse clinical settings. To address these limitations, data augmentation has emerged as a key strategy for improving model performance under limited and heterogeneous data regimes. While traditional techniques, such as affine transformations and intensity jittering, are widely used to simulate variability, they often fail to capture the structural diversity and textural realism inherent in real-world echocardiograms. Recent advances in generative models, including Generative Adversarial Networks (GANs) and diffusion models, present new opportunities for synthesizing anatomically plausible and domain-adaptive training data. This thesis introduces an augmentation framework based on conditional diffusion models specifically designed for cardiac ultrasound. We explore multiple conditioning strategies, including semantic label maps and edge maps, and assess their impact on the fidelity, diversity, and downstream segmentation of the generated images. On open-source echocardiography datasets, our proposed approach reduced distributional mismatch, as measured by a distribution-based generation metric, compared to prior GAN- and diffusion-based baselines. For segmentation tasks, synthetic-only training with our method improved mean accuracy by approximately 1--4 percentage points over earlier generative approaches, with additional but modest gains when real data were augmented with synthetic examples. Through these evaluations, the thesis provides empirical evidence that conditional diffusion models can generate clinically meaningful echocardiographic variability and improve segmentation performance, though gains remain modest compared to real-only training. These findings contribute practical insights for designing data-centric training pipelines in medical imaging and underscore the potential of diffusion-based augmentation in addressing long-standing challenges in echocardiographic machine learning applications.

Item Citations and Data

Rights

Attribution-NonCommercial-NoDerivatives 4.0 International