Rotational Cherenkov-excited luminescence scanned tomography reconstruction with symmetry vision mamba

Rotational Cherenkov-Excited Luminescence Scanned Tomography (RCELST) is an emerging optical imaging technology that visualizes the distribution of luminescent quantum yield within a treated subject. This technology involves collecting luminescence signals resulting from the excitation of luminescence probes by Cherenkov emissions induced by the rotational scanning of MV X-rays. These signals are then mapped into a sinogram for reconstructing the distribution of luminescent quantum yield by neural networks. Vision Transformers (ViTs), an effective deep learning algorithm known for capturing long-distance dependencies, have been applied to medical image reconstruction tasks. However, the large scale of medical images, combined with the quadratic complexity of ViTs, leads to erratic and time-consuming reconstruction performance. Therefore, a more efficient algorithm is essential for reducing reconstruction time while maintaining accuracy. In this study, we propose the Symmetry Vision Mamba (S-VM) to address this challenge, reducing computational time while maintaining high reconstruction accuracy. The S-VM builds on the Vision Mamba, which leverages State Space Models (SSMs) to extract global information from 2D sinogram signals. With linear computational complexity, S-VM significantly accelerates the learning process compared to the Transformer algorithms. Additionally, S-VM utilizes a symmetrical encoder architecture, incorporating convolutional stems to extract local features and enable multi-scale feature fusion by sharing parameters between the two encoder branches. Training on 10,000 sinogram signals, the S-VM algorithm achieves a peak signal-to-noise ratio (PSNR) of up to 38.67 dB and a Structural Similarity Index Measure (SSIM) of 0.97. Remarkably, these results are achieved in a Floating Point Operations (FLOPs) of 4.7G.