Fig. 1. CUDA-accelerated FPM reconstruction. (a) Batch gradient descent, where a small batch of images is randomly selected from the FPM reconstruction dataset for each gradient computation; (b) distribution of a small batch of images to CUDA thread blocks, the thread blocks are arranged in a 3D structure: the x and y directions handle the horizontal and vertical pixels of each image, while the z direction processes different images, each thread within a thread block is responsible for reading, writing, and computing a single pixel; (c) comparison between GPU parallel computing and CPU serial computing; (d) comparison of different FPM reconstruction implementations

Fig. 2. FPM imaging experiment based on cuFPM. (a) Experimental setup and optical path; (b) gradient descent using the complex-valued RMSProp optimizer; (c) working pipeline of cuFPM

Fig. 3. Reconstruction quality analysis of cuFPM on simulated data. (a) cuFPM reconstructed image; (b1)‒(b3) zoomed-in images of the low-resolution images, cuFPM reconstructed images, and the original images; (c) comparison of computation speed and reconstruction quality for different reconstruction methods

Fig. 4. cuFPM reconstruction of stained samples. (a) Fish-gill sample; (b) zoomed-in comparison of the yellow box; (c) plant-root sample; (d) zoomed-in comparison of the yellow box; (e) comparison of FPM experimental data acquisition and reconstruction time

Fig. 5. Comparison of reconstruction speed with images of different sizes