Equivariant CBCT reconstruction preprint


Medical image reconstruction with learned iterative schemes has attracted a lot of attention since the introduction of Learned Primal-Dual and Recurrent Inference Machines. However, when dealing with inherently three-dimensional problems such as Conebeam CT the computational challenges associated with primal-dual learned iterative schemes such as memory and speed become prohibitive.

In our previous work, we have shown how the associated memory limitations can be alleviated using a combination of invertible neural network blocks and patch-wise computations inside the blocks for end-to-end training. In our latest preprint, we accelerate our method and improve the parameter efficiency. Acceleration is achieved by employing a multi-scale reconstruction strategy where the initial reconstruction is performed at reduced resolution with reduced resolution of the associated primal and dual latent features. The parameter efficiency is improved by using rotationally equivariant neural networks in primal blocks, which also improves network robustness to unusual patient orientation.