A novel computational imaging autofocusing system for microscopes has been created utilizing an off-the-shelf LED and a machine learning algorithm that incorporates optical physics knowledge in its design.
Fast automatic focusing of microscopes is essential for many types of experiments. Hardware-based solutions can be fast, but they are usually expensive. And while software-based solutions can be less expensive, they usually require multiple images and thus take longer to produce accurate results.
Here, the team at UC Berkeley's Computational Imaging Lab were able to achieve the best of both worlds by designing a neural network architecture — the Fully-Connected Fourier Neural Network (FCFNN) — for fast auto-focusing that only requires the addition of a single off-axis LED as an illumination source.
Based on a single image taken under the illumination of this LED, the neural network is able to predict exactly how far to move in order to bring the sample into focus. Autofocusing occurs by taking an intensity image under single-LED illumination, Fourier transforming it, and feeding the pixels corresponding to singly-scattered light in a fully connected neural network, which predicts the movement needed to correct focus.
This technique can be applied to different types of samples, e.g. cells on converglass or a thin tissue section slice. The FCFNN architecture applies a data compression by looking only at the parts of the image that correspond to singly-scattered light. This enables it to make accurate predictions while using orders of magnitude fewer parameters than contemporary convolutional neural networks. As a result, it can be trained on a desktop computer without specialized hardware, and can be used on a similar computer or even embedded system.
The authors have provided a Jupyter notebook to fully reproduce their technique, enabling fast, low-cost autofocus in a range of microscopy applications. The research also demonstrates the performance gains that can be had by building knowledge of optical physics into machine learning systems.
Deep learning for single-shot autofocus microscopy
June 5, 2019 | Optica
Henry Pinkard, Zachary Phillips, Arman Babakhani, Daniel A. Fletcher, Laura Waller