DCT-UNet: a UNet architecture for diffuse correlation tomography

Diffuse correlation tomography (DCT) is an emerging optical technology that was developed in recent years for tissue blood flow imaging. The current methods for reconstruction of blood flow index (BFI) tomographic images from DCT signals include analytical solutions, finite element methods as well as Nth-order linear (NL) algorithm. These conventional reconstruction algorithms are based on physical principle of light electric field temporal autocorrelation functions, all of which are subjected to severely ill-posed mathematical problem that affects the accuracy and robustness of blood flow imaging. In this study, we performed a pilot attempt of deep learning framework on DCT image reconstruction, making full use of prior information and establishing the mapping relationship between optical signals and the blood flow tomographic images. For this purpose, a UNet architecture consisting of six encoders, termed as DCT-UNet network, was proposed. The first encoder is composed of convolutions at a kernel size of 1 × 1, and the second to fifth encoders are made up of deformable convolutions at a kernel size of 3 × 3. The sixth encoder is composed of standard convolutions at a kernel size of 3 × 3. Additionally, the gated units were utilized to replace MaxPooling for selection of the feature maps after convolution, retaining good features and suppressing less useful ones. The group aggregation bridge (GAB) module was adopted, rather than the ordinary skip connections, so as to link the encoder and decoder parts. The DCT-UNet model also incorporates the deep supervision from UNet++ to generate masks at different scales, which are thereafter used in the loss function and the input of the GAB. Through leveraging the strong learning capabilities of convolutional neural networks and fully utilizing the prior information of blood flow images, the DCT-UNet overcomes the primary limitations encountered with conventional methods such as ill-posed problems due to sparsity of optical measurement and tedious iterations for matrix computations, also allowing for fast blood flow imaging. Computer simulations and phantom experiments on a variety of flow anomaly configurations demonstrate the accuracy and robustness of the DCT-UNet network for DCT blood flow imaging, revealing its potential for future physiological and clinical applications.