Deep-Learning for Reduction of Deformations from Dual-Panel Breast-PET System

Dedicated dual-panel Breast-PET (B-PET) scanners have many benefits for specific breast imaging in diagnosis and treatment applications, due to their higher sensitivity, improved resolution, and lower costs compared to the whole body PET scanners. However, images from dual-panel PET systems are affected by the limited-angle artifacts due to the data angular truncations and depth-of-interaction effects for oblique lines of response (especially for objects placed close to the detector panels), leading to strong isotropic, asymmetric, and spatially variant image deformations. In this work, we investigate deep-learning based post-reconstruction deformation reduction for dual-panel TOF based B-PET imaging. While the idea presented is relevant to other neural networks as well, in this work we investigated an U-Net based architecture for artifact reduction. The network was trained on a simulated synthetic data sets providing ground truth for objects resembling experimentally acquired phantoms on which the network deformation corrections were then tested. The synthetic training data, as well as testing experimental phantom data were reconstructed using RAMLA-DIRECT reconstructions without resolution modeling, which were used as the network inputs. We demonstrate that Deep-Learning (DL) can significantly eliminate deformations of limited angle systems and improve the quantitative image quality for the B-PET system.