Zhang Yuanlong, Zhou Tiankuang, Fang Lu, Kong Lingjie, Xie Hao, Dai Qionghai
Opt Express. 2020 Jun 22;28(13):19218-19228. doi: 10.1364/OE.390878.
Wavefront sensing technique is essential in deep tissue imaging, which guides spatial light modulator to compensate wavefront distortion for better imaging quality. Recently, convolutional neural network (CNN) based sensorless wavefront sensing methods have achieved remarkable speed advantages via single-shot measurement methodology. However, the low efficiency of convolutional filters dealing with circular point-spread-function (PSF) features makes them less accurate. In this paper, we propose a conformal convolutional neural network (CCNN) that boosts the performance by pre-processing circular features into rectangular ones through conformal mapping. The proposed conformal mapping reduces the number of convolutional filters that need to describe a circular feature, thus enables the neural network to recognize PSF features more efficiently. We demonstrate our CCNN could improve the wavefront sensing accuracy over 15% compared to a traditional CNN through simulations and validate the accuracy improvement in experiments. The improved performances make the proposed method promising in high-speed deep tissue imaging.
波前传感技术在深层组织成像中至关重要，它引导空间光调制器补偿波前畸变以获得更好的成像质量。最近，基于卷积神经网络（CNN）的无传感器波前传感方法通过单次测量方法取得了显著的速度优势。然而，卷积滤波器处理圆形点扩散函数（PSF）特征的效率较低，导致其准确性欠佳。在本文中，我们提出了一种共形卷积神经网络（CCNN），通过共形映射将圆形特征预处理为矩形特征来提升性能。所提出的共形映射减少了描述圆形特征所需的卷积滤波器数量，从而使神经网络能够更高效地识别PSF特征。通过仿真，我们证明与传统CNN相比，我们的CCNN可将波前传感精度提高15%以上，并在实验中验证了精度的提升。改进后的性能使该方法在高速深层组织成像中颇具前景。
