FDA-Recon: Feature and data alignment reconstruction for sparse-view CBCT

Cone-beam computed tomography (CBCT) enables real-time three-dimensional imaging for patients, which is of great significance in improving the precision of radiotherapy and interventional procedures. Sparse-view CBCT, which can relax the readout rate of flat-panel detectors and reduce the radiation dose of X-rays, is a promising technology. However, sparse sampling can lead to streak artifacts in the reconstructed images, while the low-power X-ray source of the CBCT scanner can produce low SNR measurements. These degradations hinder accurate guidance for subsequent treatment procedures. To address these issues, this study proposes a learning-based reconstruction algorithm for sparse-view CBCT. To ensure the robustness of the proposed method in real-world scenarios, this study first constructs a large-scale simulated dataset whose distribution is close to the real data based on X-ray imaging physics and CT system characteristics, achieving coarse alignment at the data level. However, simple coarse alignment alone cannot completely bridge the gaps between simulated and real data. Therefore, this study further employs an unsupervised domain adaptation strategy to achieve deeper alignment in the feature space, ensuring the model trained on simulated data maintains its performance on real data. We refer to this strategy as FDA-Recon (Feature and Data Alignment Reconstruction). To achieve high performance in noise suppression and artifact removal, a deep neural network incorporating the novel Vision-LSTM mechanism is developed to fully exploit both local features and global dependencies in images. Results on real data from two different CBCT systems demonstrate the promising performance of the proposed image restoration neural network in artifact removal, noise suppression, and image restoration, as well as the potential of FDA-Recon in addressing sparse-view CBCT reconstruction in practical scenarios.