High-fidelity diffuse optical tomography imaging based on MRI physics information-constrained stacked autoencoder neural network

To achieve more accurate and reliable diffuse optical tomography (DOT) imaging, as well as increase the interpretability and generalizability of the DOT image reconstruction using deep learning, three distinct physics-constrained neural network (PCNN) architectures were proposed, with the stacked auto-encoder (SAE) neural network as a benchmark for comparison. These architectures directly incorporated MRI image gray values as physical prior information in three distinct ways: merging them into the network input, combining them into the loss function through a rescaling strategy by defining a total variation function, and combining both of the approaches. To investigate the effectiveness of the proposed networks, a series of numerical simulations were first performed, and the results were quantitatively evaluated and compared with the purely data-derived SAE neural network. Subsequently, the well-trained networks based on the simulation data were implemented to reconstruct the phantom experimental data to further investigate the effectiveness of the proposed methods. The experimental results revealed that the performance of the three PCNN models is superior to that of the pure neural network, with the third network architecture demonstrating the most significant advantages in terms of reconstruction fidelity and noise robustness.