Bioluminescent living organisms such as jellyfish, fungi, and fireflies possess an intrinsic mechanism of generating light through biochemical reactions. This has inspired researchers to use light as a tool for visualizing biological processes in diverse organisms, though the main challenge to working with bioluminescent proteins is the limited number of photons, or dimness of the light, released from their biochemical reaction. This has hampered the widespread use of bioluminescence for micrscopy applications. A simple way of overcoming the limitations of bioluminescent imaging is to collect light over a longer time, which enhances the signal of the image but confines the technique to processes where only very slow biological dynamics are of relevance. An innovative solution that enables improved sensitivity and speed of bioluminescence imaging, which enhances signal-to-noise, is required for further advancement of the technique.
The key solution proposed by Morales-Curiel et al (2022) is the combination of bioluminescent microscopy with computational methods such as artificial neural networks (NN). Just as the human brain can recognize a familiar voice among a noisy crowd, NNs simulate the behavior of the human brain by automatically extracting relevant features in bioluminescence imaging data and weighting the importance of each to classify and group that data. In this article, we describe a protocol to train a NN to perform bioluminescence denoising and three-dimensional reconstruction to improve the speed and the quality of the images collected from bioluminescent samples.
First, we demonstrate that it is possible to train a NN to denoise bioluminescence images and use it to denoise different biological samples like cells grown in culture, the muscles of the nematode worm C. elegans, clusters of C. Elegans neurons, as well as in zebrafish embryos. To accomplish this, we first collected a dataset of noisy, or dim, images and paired them with their high-quality counterparts, where the signal-to-noise ratio was significantly higher. Next, we trained the NN to map noisy images to high-quality images. The learning process is done by iteratively comparing the response of the NN against its high-quality counterpart, if the prediction of the network is different from the given high-quality image, the internal parameters of the NN are tweaked so the next iteration generates an image closer to the target.
Second, we employed a microscopy technique called light field imaging, which allows the acquisition of the structure of a three-dimensional object from a single picture. Similar to how a fly´s eyes work, instead of using one lens to focus the image from the bioluminescent sample onto the camera, light field microscopy works by placing several micro-lenses in an array to acquire different perspectives of the bioluminescent sample. Next, we specifically trained a NN to interpret this segmented image and generate a volumetric view of the sample. To accomplish this, the NN is presented with a dataset containing single images taken with the micro-lenses and mapped to their corresponding three-dimensional views. However, due to the low amount of light coming from the bioluminescent sample, the three-dimensional reconstruction obtained from the single bioluminescent image was blurry due to the high amount of noise in the background. As a workaround, we decided to couple the denoising network and the 3D reconstruction network improve the network inference. Surprisingly, this new pipeline achieved impressive results by increasing the quality of the reconstruction and the results were very close to the expected targets.
Third, we tested the efficacy of their system in recording fast bioluminescent processes, such as calcium dynamics in moving samples, namely in the nematode worm C. elegans. Calcium dynamics in the body wall muscles of the C. elegans is well characterized due to its role in the muscle contraction that generates movement. Therefore, as the worm crawls, a higher calcium signal is expected in contracting muscles than in relaxed muscles. After the processing of the raw bioluminescent images using the pipeline described above, the authors observed the largest intensity in calcium signal mapped to positive body curvatures, which can be explained by muscle contraction and calcium influx during this motion.
Overall, the use of computational algorithms such as NNs can push the limitations of microscopy techniques such as bioluminescence microscopy and thus present an appealing alternative to the common fluorescence microscopy techniques used in most research. In the future, with the generation of more efficient bioluminescent proteins coupled with the design of simpler, more efficient NN architectures, the timing and quality of the acquired bioluminescent data are bound to significantly improve.
Morales-Curiel, 2022 [Luis Felipe Morales-Curiel et al. Volumetric imaging of fast cellular dynamics with deep learning enhanced bioluminescence microscopy. Communications Biology. 2022]