Training End-to-End Unrolled Iterative Neural Networks for SPECT Image Reconstruction

Training end-to-end unrolled iterative neural networks for single-photon emission computerized tomography (SPECT) image reconstruction requires a memory-efficient forward-backward projector for efficient backpropagation. This article describes an open-source, high-performance Julia implementation of a SPECT forward-backward projector that supports memory-efficient backpropagation with an exact adjoint. Our Julia projector uses only similar to 5% of the memory of an existing MATLAB-based projector. We compare unrolling a CNN-regularized expectation-maximization (EM) algorithm with end-to-end training using our Julia projector with other training methods, such as gradient truncation (ignoring gradients involving the projector) and sequential training, using XCAT phantoms and virtual patient (VP) phantoms generated from SIMIND Monte Carlo (MC) simulations. Simulation results with two different radionuclides (Y-90 and Lu-177) show that: 1) for Lu-177 XCAT phantoms and 2) Y-90 VP phantoms, training unrolled EM algorithm in an end-to-end fashion with our Julia projector yields the best reconstruction quality compared to other training methods and ordered-subset EM (OSEM), both qualitatively and quantitatively. For VP phantoms with Lu-177 radionuclide, the reconstructed images using end-to-end training are in higher quality than using sequential training and OSEM, but are comparable with using gradient truncation. We also find there exists a tradeoff between computational cost and reconstruction accuracy for different training methods. End-to-end training has the highest accuracy because the correct gradient is used in backpropagation; sequential training yields worse reconstruction accuracy, but is significantly faster and uses much less memory.

Automated summary

This article introduces an end-to-end unrolled iterative neural network training method for single-photon emission computerized tomography (SPECT) image reconstruction, which requires a memory-efficient forward-backward projector. The authors present an open-source Julia implementation of this projector, which uses only a fraction of the memory compared to a MATLAB-based projector. The study compares the performance of the proposed training method with other approaches using various phantoms and demonstrates that the end-to-end training with the Julia projector achieves the best reconstruction quality.