Joint Rigid Motion Correction and Sparse-View CT via Self-Calibrating Neural Field
share
Neural Radiance Field (NeRF) has widely received attention in Sparse-View (SV) CT reconstruction problems as a self-supervised deep learning framework. NeRF-based SVCT methods model the desired CT image as a continuous function that maps coordinates to intensities and then train a Multi-Layer Perceptron (MLP) to learn the function by minimizing loss on the SV measurement. Thanks to the continuous representation provided by NeRF, the function can be approximated well and thus the high-quality CT image is reconstructed. However, existing NeRF-based SVCT methods strictly suppose there is completely no relative motion during the CT acquisition because they require accurate projection poses to simulate the X-rays that scan the SV sinogram. Therefore, these methods suffer from severe performance drops for real SVCT imaging with motion. To this end, this work proposes a self-calibrating neural field that recovers the artifacts-free image from the rigid motion-corrupted SV measurement without using any external data. Specifically, we parametrize the coarse projection poses caused by rigid motion as trainable variables and then jointly optimize these variables and the MLP. We perform numerical experiments on a public COVID-19 CT dataset. The results indicate that our model significantly outperforms two latest NeRF-based methods for SVCT reconstruction with four different levels of rigid motion.
READ FULL TEXT