Cloud-Trained Deep Learning for Microscopic Image Reconstruction

The patent divides the large-scale imaging task into multiple smaller image acquisitions that are subsequently reconstructed using deep learning algorithms. The imaging area is segmented into multiple fields of view that are captured sequentially at high speed, then computationally reassembled to create a large-scale image without requiring slow point-by-point scanning across the entire area.

The patent performs preliminary image reconstruction and enhancement using trained deep learning networks before final image assembly. The network is pre-trained on large datasets to learn reconstruction patterns, allowing rapid processing of acquired images while maintaining high quality output, thus speeding up the overall imaging workflow.

The patent introduces a deep learning reconstruction network as an intermediary between the raw acquired images and the final output. This network learns to reconstruct high-quality images from low-signal input data, effectively mediating between the need for fast scanning and the requirement for high signal-to-noise ratio without requiring slow scanning speeds.

The patent changes the processing parameters through learned transformations in the deep learning network. The network learns optimal parameter transformations during training that enhance signal-to-noise ratio in the reconstructed images while maintaining the fast scan speeds used during data acquisition, effectively decoupling scan speed from output quality.

The patent creates a universal deep learning network that can be deployed across multiple microscopic systems. The network is trained on diverse data from various microscope types and configurations, making it universally applicable. This single trained network serves multiple systems simultaneously, eliminating the need for each system to have its own separately trained network, thus reducing overall complexity while maintaining high processing quality.

The patent creates a master trained network model that can be copied and deployed to multiple microscopic systems. Rather than training separate networks for each system, the trained weights and parameters are replicated across different devices, providing consistent high-quality image processing without repeating the computationally intensive training process at each location.

The patent extracts and separates the training phase from the inference phase. Large datasets are transferred to the cloud only once for the computationally intensive training process, while the resulting trained network model (which is much smaller) is then deployed locally for rapid inference. This extraction of the training function to the cloud maximizes processing power utilization without requiring continuous data transport.

The patent performs the computationally intensive network training in advance using cloud computing resources. The trained network model is then stored and reused for multiple inference tasks. This preliminary action of training once in the cloud eliminates the need for repeated data transfers, as the same trained model can process numerous images locally without further cloud communication.