Fig. 3: Reconstruction with less data.
From: Neural network-based processing and reconstruction of compromised biophotonic image data
a Schematic representation of a typical reconstruction process. It consists of a dense set of input data and a standard algorithm. b With deep learning, a significant reduction from the traditional multi-layered stack is achieved for image reconstruction. This input is then fed into a neural network, which interprets and reconstructs the data. c Optimization of LED configuration using deep learning for Fourier ptychography with the resulting amplitude and phase components. An example from the evaluation dataset is provided for comparison, showcasing the phase component of the iterative Fourier ptychography reconstruction, which serves as the ground truth, alongside the output of the neural network, together with a cross-sectional analysis15 [figure adapted with permission from ref # 15 © Optical Society of America]. d The Recurrent-MZ volumetric imaging framework is illustrated through examples of 3D imaging of C. elegans, showcasing the initial input scans, the output processed by the network, and the established ground truths for comparison12. e The SS-OCT system acquires raw OCT fringes, from which the target image of the network is derived by directly reconstructing the original OCT fringes. By processing an undersampled image through a trained network model, an OCT image free of aliasing is produced, closely aligning with the ground truth. The provided example involves a 2× undersampled OCT image14. f Following a swift process of transfer learning, the RNN few shot hologram model demonstrates excellent generalization capabilities on test slides of new types of samples (lung tissue sections)
