Spatially Variant Point Spread Functions Deconvolution with Untrained Neural Network for Reflection Ultrasound Computed Tomography Imaging

Ultrasound computed tomography (UCT) enables tissue characterization through both reflection and transmission imaging. In reflection UCT imaging, the point spread function (PSF) varies spatially due to system limitations, necessitating spatially variant PSF deconvolution for high-quality image reconstruction. Traditional methods typically rely on explicit priors such as Gaussian or sparsity constraints, which limit their generalization. To overcome this, we propose a spatially variant PSF deconvolution algorithm based on an untrained neural network (UNN), which employs implicit regularization to enhance resolution and suppress artifacts without requiring training data. Our approach is grounded in the UCT image degradation model and leverages a UNN to recover the underlying image. Spatially variant PSFs were experimentally measured using a line target and interpolated into a dense, pixel-wise PSF map via Kringing interpolation. The proposed method was evaluated on UCT data from the human arm. Compared with traditional deconvolution algorithms based on Gaussian and sparsity priors, our approach achieved improvements of 22.35% and 15.26% in terms of the Brenner gradient metric, respectively. Furthermore, intensity profile analysis confirmed that the proposed method yields clearer and sharper image reconstructions.