Abstract
Circular-Scan photoacoustic tomography (PAT) can provide high-resolution images of optical absorption, but its analytical reconstructions, such as delay-and-sum (DAS), are highly sensitive to scanning radius (SR) inaccuracies, which cause severe geometric distortions and artifacts. In this work, we propose a deep learning framework, termed smooth deconvolution ResNet (SD-ResNet), to correct DAS reconstruction degradation induced by SR errors. SD-ResNet uses an ImageNet-pretrained ResNet-50 encoder and a lightweight deconvolutional decoder with additional smoothing convolutions to suppress checkerboard artifacts and restore fine structural details. A paired training dataset is generated using k-Wave simulations driven by human thoracic computed tomography (CT) slices: for each phantom, radiofrequency data are simulated once, and DAS images reconstructed with the true SR serve as ground truth, whereas images reconstructed with biased SR values serve as inputs. This design provides structurally diverse training samples and enhances generalization. In silico experiments show that SD-ResNet effectively recovers image quality across a range of SR deviations. Phantom experiments with polyethylene microspheres further confirm that the proposed method can substantially reduce artifacts and recover correct source shapes under practical SR mismatches, offering a robust tool for SR-error-resilient PAT imaging.