Physical Parameter-Guided Deep Learning Ultrasound Localization Microscopy Framework Based on Diffusion Model.
Authors
Abstract
Deep learning based-imaging methods have demonstrated significant potential for achieving high spatiotemporal resolution in Ultrasound Localization Microscopy (ULM), a state-of-the-art hemodynamic microvascular imaging method. The accurate application of deep learning in ULM remains substantially constrained by the scarcity of reliable ground truth data and diverse training datasets across varying imaging conditions. However, quantitative microbubble imaging data remain unavailable in society, posing a significant challenge for deep learning-based approaches. To address this critical data gap, we propose, for the first time, a Physical parameter-guided Diffusion (PgD) framework capable of synthesizing microbubble images under diverse imaging conditions by incorporating transducer specifications and acoustic waveform parameters. These synthetic images serve as a robust training dataset for deep learning-based ULM, enabling accurate model development despite the lack of experimental ground truth. Our systematic evaluation across hundreds of imaging parameter sets demonstrates that the synthetic microbubble data generated by our framework exhibit high fidelity to experimental ground truth, with an average Structural Similarity Index Measure reaching 0.97. When applied to deep learning-based ULM training, the proposed framework outperforms conventional ULM methods and alternative training data strategies, achieving 5-10 μm improvement in spatial resolution while requiring fewer frames for accurate reconstruction across diverse imaging scenarios. The consistent performance improvement validates the robustness and generalizability of our PgD framework, demonstrating its significant value for advancing deep learning in ULM applications.