Zhao Xudong, Hu Shuguo, Yang Qiang, Zhang Zhiwei, Guo Qianjin, Niu Chaojun
Academy of Artificial Intelligence, Beijing Institute of Petrochemical Technology, Beijing, 102617, China.
College of Computer Science and Technology, Inner Mongolia University, Hohhot, 010021, China.
Photoacoustics. 2025 Feb 8;42:100695. doi: 10.1016/j.pacs.2025.100695. eCollection 2025 Apr.
Photoacoustic tomography (PAT) provides high-contrast, high-resolution biomedical images at rapid speeds. However, the quality of these images is highly sensitive to sampling density. Sparse sampling can significantly reduce equipment costs but often leads to image artifacts and degraded quality. While deep learning models have greatly enhanced sparse PAT imaging, their high computational requirements limit their use in resource-constrained environments. To overcome this challenge, we propose AD-WaveNet, a lightweight network that integrates the Discrete 2D Wavelet Transform (DWT) with adaptive attention mechanisms. This approach enhances sparse image reconstruction while reducing computational complexity. The attention mechanisms are specifically designed to exploit the multi-scale decomposition properties of DWT, allowing the model to emphasize key features across various scales. Compared to the latest models, AD-WaveNet reduces computational complexity and parameter count by nearly two orders of magnitude, while maintaining optimal reconstruction quality. This demonstrates AD-WaveNet's significant potential for practical applications in PAT imaging.
光声层析成像（PAT）能够快速提供高对比度、高分辨率的生物医学图像。然而，这些图像的质量对采样密度高度敏感。稀疏采样可以显著降低设备成本，但往往会导致图像伪影和质量下降。虽然深度学习模型极大地增强了稀疏PAT成像，但它们对计算的高要求限制了其在资源受限环境中的应用。为了克服这一挑战，我们提出了AD-WaveNet，这是一种将离散二维小波变换（DWT）与自适应注意力机制相结合的轻量级网络。这种方法在降低计算复杂度的同时增强了稀疏图像重建。注意力机制经过专门设计，以利用DWT的多尺度分解特性，使模型能够在不同尺度上强调关键特征。与最新模型相比，AD-WaveNet将计算复杂度和参数数量降低了近两个数量级，同时保持了最佳的重建质量。这证明了AD-WaveNet在PAT成像实际应用中的巨大潜力。
