AI-powered deep learning framework restores high-quality optoacoustic images with 87.5% fewer detectors

A new deep learning framework developed by researchers at Fudan University enables high-fidelity multispectral optoacoustic tomography (MSOT) using only 32 detectors—an 87.5% reduction from the standard 256-detector configuration—while maintaining image quality comparable to conventional full-view systems. The work, published in BME Frontiers, could significantly lower the hardware cost and complexity of optoacoustic imaging, accelerating its translation to clinical practice.

MSOT combines optical and ultrasound technologies to provide anatomical, functional, and molecular information with deep tissue penetration and high spatial resolution. However, clinical translation has been hampered by the need for dense sensor arrays and full-view acquisition, which increase hardware complexity, cost, and reconstruction time. Sparse-view imaging has emerged as a solution, but the resulting data insufficiency introduces severe noise, aliasing artifacts, and reduced resolution.

The team developed OA-UDNet (Optoacoustic Universal Denoising Network), a hybrid diffusion-based framework that performs joint denoising and high-fidelity image restoration within a unified generative model. Unlike conventional methods that treat these tasks separately, OA-UDNet integrates a short-chain diffusion process with an edge-aware self-attention module that preserves anatomical boundaries using fixed Sobel priors. The framework uses only 10 diffusion steps—drastically faster than conventional diffusion models that require hundreds to thousands of steps—achieving inference speeds of approximately 150 ms per frame.

Trained on more than 250,000 in vivo images, OA-UDNet consistently elevated PSNR from baseline levels of ~20–23 dB to ~31–39 dB across brain, abdomen, hindlimb, and tumor datasets, representing relative improvements of 62% to 77%. Structural fidelity was restored from ~0.72–0.81 to near-ideal values of 0.96–0.98. The framework also demonstrated zero-shot generalization on human clinical data, successfully reconstructing images of human calf vasculature without any human-specific training.

The method substantially reduces functional spectral unmixing errors—under the 32-detector configuration, normalized root mean square error for oxygenated hemoglobin decreased from 7.97% to 1.56%—ensuring that functional imaging capabilities are preserved even with sparse data.

"OA-UDNet establishes a new benchmark for sparse-sampling MSOT," said Dean Ta, corresponding author and professor at Fudan University. "This approach considerably alleviates the hardware density requirements of current optoacoustic systems, offering a highly reliable and cost-effective computational pathway toward clinical translation of high-fidelity functional imaging."