Diffusion-based model as a generalizable denoiser for optoacoustic imaging

The growing use of optoacoustic imaging in preclinical and clinical studies has fostered the recent development of cost-effective systems facilitating wider accessibility of this technology to researchers and physicians, particularly in resource-limited settings. Low-cost light sources have been considered for cost reduction, but the resulting images are commonly afflicted with different types of artifacts and noise. Advanced denoising methods thus become essential for optimizing imaging performance and maximizing the diagnostic potential of optoacoustic systems. In this work, we introduce a novel denoising framework based on improved denoising diffusion probabilistic models (DDPMs) tailored to address heterogeneous noise sources inherent to different acquisition systems and experimental conditions. Unlike traditional diffusion-based models, the proposed model employs Gaussian noise as a carrier to suppress both Gaussian and complex non-Gaussian distortions while effectively preserving critical image details. We devised an iterative mechanism to gradually dispose of noise artifacts patterns with the backward process of the denoising-enhanced DDPM model. Experimental results using laser-diode and light-emitting-diode sources demonstrate the proposed method's consistent performance improvement over state-of-the-art and benchmark approaches like block-matching and 3D filtering, deep image prior, and improved DDPM.