DGDOT-Net: A Deep Generative Model With Attention Fusion for Enhanced High-Density Diffuse Optical Tomography

Abstract

Functional near-infrared spectroscopy (fNIRS) noninvasively evaluates the optical properties of target tissues to monitor functional changes. High-density diffuse optical tomography (HD-DOT) based on this technology enables high-resolution 3-D reconstruction. However, the strong scattering of photons by brain tissue limits the ability of detected signals to accurately reflect changes in brain function, reducing both the accuracy and 3-D resolution of fNIRS-based reconstructions. This article introduces a deep generative model, DGDOT-Net, which incorporates an attention fusion mechanism to enhance the imaging resolution and robustness. The model first decouples key features in the inverse mapping process between observed signals and reconstructed results, leveraging the conditional variational autoencoder (CVAE) architecture to model the probability distribution in latent space and regulate the reconstruction outcome. In addition, a depth-aware attention mechanism embedded within the encoder and decoder extracts effective features from the progressive encoding process, improving learning efficiency. This study first demonstrates the superior reconstruction performance of the model through a series of numerical simulation experiments and evaluates its robustness under low signal-to-noise ratios and varying medium conditions. Specifically, the average values of structural similarity index (SSIM), peak signal-to-noise ratio (PSNR), mean absolute error (MAE), contrast-to-noise ratio (CNR), R, and Jaccard Index achieved by DGDOT-Net on simulated data are 0.83, 21.03 dB, $2.75\times 10 ^{\mathrm {-3}}$ , 7.10, 0.45, and 0.73, respectively. Subsequently, physical phantom data collected using a locally developed prototype system are tested, yielding average metric values of 0.87, 18.65 dB, $15.32\times 10 ^{\mathrm {-3}}$ , 9.61, 0.79, and 0.92, respectively. Furthermore, DGDOT-Net demonstrates the ability to reconstruct optical properties at a depth of 1.5 cm with a spatial resolution of 1 cm. The experimental results confirm that the proposed model enhances the 3-D reconstruction of brain functions using fNIRS, advancing the clinical applicability of related technologies.