Conditional Sampling of High-Quality Ultrasound Images from Diffusion Models with Limited Data

Gurunath Reddy, Sushanth Govinahallisathyanarayana, Prasad Sudhakara Murthy, Chandan Aladahalli, Dattesh D. Shanbhag

Keywords: Image quality, Diffusion models, Ultrasound, Synthetic data

Registration Requirement: Yes

Abstract: Diffusion models can synthesize realistic medical images but are computationally intensive, requiring thousands of sampling steps. Additionally, since the conventional sampling process is stochastic, it can produce low-quality or irrelevant outputs, especially when trained with less data. In this work, we introduce a quality-aware sampling strategy that monitors image fidelity during generation, terminating trajectories that are likely to lead to low quality images early, and halting the process once satisfactory quality is achieved. This approach accelerates diffusion-based synthesis, decreases computational overhead, and yields anatomically plausible images that support the development of more robust healthcare AI models.

Visa & Travel: Yes

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 53