Deep Learning-Based Denoising Model Boosts Microscope Imaging Speed by 16 Times

2020-01-19

To help speed up microscope systems, researchers from the Salk Institute at the University of Texas at Austin have developed a new AI-based microscopy imaging method that is expected to be 16 times faster than the microscope technology used for brain imaging.


In a paper titled "Deep Learning-Based Point-Scanning Super-Resolution Imaging" published on bioRxiv, the researchers pointed out that "point-scanning imaging systems are perhaps the most widely used tools for high-resolution cellular and tissue imaging. Like all other imaging modalities, it is challenging to simultaneously optimize the resolution, speed, sample preservation, and signal-to-noise ratio of point-scanning systems." To address this challenge, Uri Manor from the Salk Institute started using an image processing technique called deconvolution, which astronomers have used to achieve higher resolution in telescope images of stars and planets.


After collecting enough data to build a training dataset containing high-resolution and super-resolution images, the team trained the model using a ResNet-based U-Net convolutional neural network.


The researchers explained in the paper: "To train the model for this purpose, it is necessary to have many perfectly paired high-resolution and low-resolution image pairs. We chose to generate high-resolution images synthetically to simulate low-resolution image pairs acquired under the microscope, rather than manually acquiring high-resolution and low-resolution image pairs for training, to generate semi-synthetic training data."


After training the model, the team tested it and applied it to images created using different microscopes in other laboratories.


Manor said, "Typically in deep learning, you have to retrain and fine-tune models for different datasets, but we're excited that our system works so well across a variety of samples and image sets." Manor and the team hope to develop software that can reconstruct images in real-time, allowing researchers to immediately see super-resolution images.


In the paper, the team explained, "It is worth noting that our EM PSSR model can recover undersampled images obtained through different optical devices, detectors, samples, or sample preparation methods in other laboratories. We demonstrate that undersampled confocal images combined with multi-frame PSSR models trained on Airyscan delays have approximately 100 times lower laser dose and 16 times higher frame rate compared to the corresponding high-resolution acquisition, contributing to Airyscan-equivalent spatial resolution and SNR. In summary, PSSR enables acquisition of point-scanning images with resolution, speed, and sensitivity that cannot be achieved otherwise."


The team has released the corresponding code on GitHub.

RECENT AI NEWS

RECENT AI TOOLS