|
Photoacoustic imaging (PAI) is an emerging modality that has generated increasing interest for its uses in clinical research and translation. To fully exploit its potential for various preclinical and clinical applications, it is necessary to develop systems that offer high imaging speed, reasonable cost, and manageable data flow. Currently, a significant challenge lies in the fabrication of ultrasound arrays, as many of them are not densely populated enough to fully sample the signals. Ideally, the pitch of the arrays should be half of the center ultrasonic wavelength to prevent spatial undersampling and the subsequent reconstruction artifacts, such as aliasing artifacts and structural deformation. Here, a novel photoacoustic sparse sampling Transformer-CNN coupling network (passFormer) is proposed to decouple target details and spatial under-sampling artifacts from high-frequency image information in a heterogeneous feature-aware manner. To be specific, we first decompose the sparse sampling (SS) photoacoustic (PA) images into 2 parts: high-frequency (HF) and low-frequency (LF) compositions. Our methodology incorporates two bridging modules, the LF modules and the HF modules. The LF coupling module extracts content features (Xc) and latent texture feature (Xtex), and the HF coupling module extracts high-frequency embedding (Xemb) containing the target details features (Xdetail) and under-sampling artifacts (Xart). We feed Xt and Xemb into a modified transformer with three encoders and decoders to obtain well-refined HF texture features. At last, we combine the refined HF texture features with pre-extracted Xc by pixel-wise summation reconstruction. Experimental results on publicly available full and sparse reconstruction datasets of mouse and phantom PA images highlight the superior performance of our method, particularly in live mouse imaging. This new approach enables accelerated data acquisition and image reconstruction, facilitating the development of practical and cost-effective imaging systems.
|
