LoGoReg-Net: local-global feature aggregation regularized unrolling network for functional brain imaging via high-density DOT.

Functional near-infrared spectroscopy (fNIRS) is a non-invasive brain imaging technique that is widely utilized in clinical and rehabilitation settings. High-density diffuse optical tomography (HD-DOT) enhances the spatial resolution of fNIRS by densely arranging light sources and detectors. However, due to the ill-posed nature of HD-DOT, achieving high reconstruction accuracy and computational efficiency remains challenging. Unrolling networks offers a promising solution by combining the advantages of traditional optimization algorithms and deep learning. Nevertheless, existing methods are still limited in effectively modeling both local and global characteristics of optical perturbations and often incur high computational costs, hindering their applicability. To overcome these limitations, this paper proposes an unrolling network named LoGoReg-Net, which incorporates a local-global feature aggregation module (LGFAM) as a learnable regularization term. LGFAM consists of two branches: a multi-scale local feature capture module, which enhances the representation of local features at multiple scales, and a triple-coordinate global feature aggregation module, which efficiently models the spatial distribution of global optical perturbations. The effectiveness of the proposed method is validated through both numerical simulations and physical phantom experiments. The results demonstrate that LoGoReg-Net consistently outperforms existing approaches on in-distribution, out-of-distribution, and real-world datasets in terms of SSIM and PSNR. It further shows superior structural fidelity, fine-detail recovery, and robustness, with these performance gains consistently preserved when extended to data acquired at different wavelengths. This approach provides an effective solution to overcome current bottlenecks in HD-DOT imaging and holds significant promise for advancing high-performance brain functional imaging technologies toward clinical translation.