Fig. 1. CUDA method of Cu-3DSIM. (a) Method of 3DSIM initialization and the reconstruction process. (b) Parallel GPU acceleration method of cross-correlation algorithms. (c) Separation matrices for the GPU parallel method.

Fig. 2. Optimization of the reconstruction algorithm for the GPU method. (a) Traditional method of 3DSIM shifting and the notch filter process, and (b) the shifting and notch filter process in the proposed Cu-3DSIM method. (c) Calculation of the space occupation of the core variables in the Open-3DSIM reconstruction process, where the intermediate variables are not included in the counted variables. (d) Calculation of the space occupation of the core variables in the Cu-3DSIM reconstruction process, where the intermediate variables are not included in the counted variables.

Fig. 3. Comparison of reconstruction time and memory allocation. (a) Reconstruction time of 512×512 image size samples. (b) Reconstruction time of 1024×1024 image size samples. (c) Reconstruction memory space of 1024×1024 image size samples.

Fig. 4. Complex background sample comparison experiment. (a) is the reconstruction result based on the Cu-3DSIM method, (b) is the reconstruction result based on the Open-3DSIM method, and (c) is the reconstruction result based on the Open-3DSIM with the notch filter method.

Fig. 5. Cu-3DSIM performance compared with other methods. (a)–(e) are the reconstruction results based on different methods. (f), (g) are the quantitative evaluations of the reconstruction intensity and SNR. (h) is the reconstruction results of low SNR samples. (i) is the max intensity projections of the reconstruction results.

Fig. 6. Long-term reconstruction experiment. (a) is the reconstruction results based on the Cu-3DSIM method, and (b) is the reconstruction results based on the Open-3DSIM method.