Point-Cloud Transformer for 3-D Electrical Impedance Tomography

Electrical impedance tomography (EIT) is an emerging medical imaging modality that offers nonintrusive, label-free, fast, and portable features. However, the three-dimensional (3-D) EIT image reconstruction problem is thwarted by its high dimensionality and nonlinearity, thus suffering from low image quality. This article proposes a novel algorithm named point-cloud transformer for 3-D EIT image reconstruction (ptEIT) to tackle the challenges of 3-D EIT image reconstruction. ptEIT leverages the nonlinear representation ability of deep learning and effectively addresses the computational cost issue by using irregular-grid representation of the 3-D conductivity distribution in point clouds. The permutation invariant property rooted in the self-attention operator makes ptEIT particularly suitable for processing this type of data, and the objectwise chamfer distance (OWCD) effectively solves the mean-shaped behavior problem encountered in reconstructing multiple objects. Our experimental results demonstrate that ptEIT can simultaneously achieve high accuracy, spatial resolution, and visual quality, outperforming the state-of-the-art 3-D EIT image reconstruction approaches. ptEIT also offers the unique feature of variable resolution and demonstrates strong generalization ability toward different noise levels, showing evident superiority over voxel-based 3-D EIT approaches.