Multi-scale wavelet network for high-quality sparse photoacoustic imaging

This study introduces the Wavelet-Based Sparse Photoacoustic Reconstruction Network (WSPARN), designed to enhance the reconstruction quality of photoacoustic images under sparse sampling conditions. WSPARN utilizes the discrete wavelet transform (DWT) for multi-scale feature extraction, replacing traditional pooling layers with DWT-based down-sampling. The network applies channel attention to low-frequency components and spatial attention to high-frequency components to enhance image quality. A dynamic loss function that combines mean squared error (MSE), edge, and frequency domain losses supports optimal model convergence. Validation on simulated vascular imaging data and real mouse photoacoustic tomography scans demonstrated that WSPARN not only significantly reduces artifacts and better preserves structural details compared to recent models from the past few years, but also offers computational advantages. A dedicated section on Complexity Comparison and Analysis highlights these computational benefits, showing WSPARN’s lower computational demand compared to other advanced models. Finally, an ablation study confirmed the benefits of wavelet-based down-sampling in improving reconstruction accuracy, substantiating the model’s effectiveness in practical applications.