Artificial Intelligence (AI) and deep learning methods are making seemingly impossible tasks now possible. From recovering contrast to improving signal-to-noise ratio, or for new approaches to managing challenging acquisition parameters or segmentation previously difficult or nearly impossible, these approaches can now be automated thanks to AI.
The NIS-Elements NIS.ai suite consists of various modules and functions which expand the NIS-Elements platform by building in tailor-made solutions for acquisition, visualization and analysis.
Intensity measurements were desired to be made along the nuclear envelope of cells. Conventional segmentation could not differentiate the cellular structures and misses several cells. AI-trained segmentation recognizes and identifies the nuclear envelope successfully.
Os dados de fluorescência de campo amplo podem ser contaminados por luz dispersa e fora de foco, mas as ferramentas baseadas em AI podem recuperar o contraste da imagem, removendo o ruído e o desfoque.
Clarify.ai uses artificial intelligence to automatically remove blur from fluorescence microscope images.
Clarify.ai utilizes new Nikon technologies executed on graphic processing units (GPUs) to leverage fast and efficient clarity in images normally corrupted by blur due to out-of-focus light.
The module is pre-trained to recognize fluorescence signal emitted from out-of-focus planes and can computationally remove this haze component from the image automatically, leaving behind the in-focus structures, and can be used on any widefield 2D or 3D data set, detector, or magnification, without the need for AI training or introduction of bias from complicated user-settings.
60x Multichannel 3D widefield Z stack before and after application of Clarify.ai
20x Multichannel 2D widefield image of a scattering tissue slice before and after application of Clarify.ai
The NIS.ai processing and analysis module consists of tools dedicated to improving and enhancing efficiency in data acquisition and simplifying previously complex or difficult analysis routines.
By recognizing patterns present in two different imaging channels, Convert.ai can be trained to predict what the second channel would look like when only the first channel is acquired.
Commonly, this can be used as a segmentation tool for label-free approaches, or imaging without harmful near-UV excitation. Once the neural network learns the pattern common to two channels, then in subsequent experiments the second channel is no longer needed to be acquired. Throughput of acquisition as well as specimen viability both increase as a result.
A coloração de núcleos com DAPI é um método comum que permite a contagem e segmentação de células. O Convert.ai pode ser treinado para prever onde o rótulo DAPI está presente nas imagens DIC ou de fase. Este canal previsto pode, então, ser usado para segmentar e contar, sem a necessidade de marcar a amostra com DAPI ou adquirir um canal de fluorescência.
Fotos cortesia do Dr. Kentaro Kobayashi, Divisão Técnica, Instituto de Pesquisa de Ciência Eletrônica, Universidade de Hokkaido
Some fluorescent samples express a very low signal and it is difficult to visualize or extract details for segmentation.
In addition, many of these samples are sensitive to light or photobleach very quickly and need to be imaged as fast as possible.
Enhance.ai can restore details by training the network what properly-exposed images look like. Then this recipe can be applied to underexposed images to restore detail that can be used for further analysis.
DAPI stained nuclei are purposefully underexposed to limit the specimen’s exposure to near-UV light. Enhance.ai is used to restore the signal-to-noise ratio to normally exposed DAPI staining, for easy segmentation and counting.
Some images are nearly impossible to segment by traditional intensity thresholding methods. A neural network can be trained by human classification of structures of interest that cannot easily be defined by classic thresholding and image processing by using Segment.ai.
By tracing features of interest and training these compared to the underlying image, the neural network can learn and apply segmentation to similar images, recognizing features previously only identifiable by painstaking manual tracing approaches.
Por meio de contraste de fase, não foi possível definir os processos neuronais com precisão pelo limiar tradicional. O Segment.ai foi treinado em processos neuronais identificados à mão (reconhecidos por humanos), aprendendo a rastreá-los em imagens subsequentes.
Included in the NIS-Elements AR core package, Denoise.ai can be applied to confocal images to remove shot noise. All images contain shot noise, which is a Poisson-distributed noise related to discreetly sampling (acquiring images) of a continuous event. As signal levels decrease, the contribution of shot noise increases and noisy images result, following a square-root function. Such noise therefore is modeled in a neural network and doesn’t need to be further trained.
With new fluorescent techniques pushing intensities lower and acquisition speeds increasing, Denoise.ai can recognize and remove the shot noise component of images, increasing clarity and allowing for shorter exposure times or more exposures of specimens while maintaining viability.
O Denoise.ai pode ser aplicado para remover o componente de ruído de disparo das imagens, sem alterar a estrutura por trás da imagem e os valores de intensidade.
Clarify.ai and Denoise.ai are pre-trained deep learning networks and require no additional settings or parameters in order to apply these tools automatically to images.
The NIS.ai Processing and Analysis module uses training data to specifically target and address user-defined experiment parameters: it employs convolutional neural networks (CNNs) to learn from labeled training data created by either conventional segmentation or human-assisted tracing of small subset of representative samples.
When using the module, the software interface makes it easy to apply complex deep learning to sample data, eliminating the need to design a complex neural network and apply training data to it.
Automated tools take this training data and apply the neural network to recognize patterns. The result training recipe can then be applied repeatedly and reliably to similar samples to process or analyze huge volumes of data at significantly faster speed than traditional techniques.
The NIS-Elements GA (General Analysis)/GA3 option enables easy customization of complex analysis or statistical flows such as 3D volume measurement and 4D tracking by simply dragging and dropping analysis templates, ensuring accurate and reliable analyses.
General Analysis is used to apply Convert.ai to brightfield images to mark nuclei, and Denoise.ai applied to a noisy fluorescence channels. The converted channels can then be tracked over time lapse to measure cell movements. This routine is then applied to multiple data sets for data measurement.
By combining functions of NIS-Elements such as a General Analysis processing toolbox (GA/GA3) and acquisition workflow toolbox (JOBS) for customizing complex experiments, a user can develop various image acquisition protocols and streamline the entire experimental workflow, from image acquisition to analysis.
Various functions performed by applying NIS.ai such as detection and analysis of specific cell states can be incorporated into image acquisition sequences that influence the control parameters of the acquisition device during an experiment based on feedback from analysis results, improving throughput and enabling the building of more complex experimental systems.
Example of incorporating GA and NIS.ai into JOBS
The image above shows an example of Segment.ai being used in an experiment. After multipoint imaging, AI is used to detect target cells. If target cells are detected, the result is fed back into the experimental sequence, and light stimulation or changes to the imaging conditions are performed. If target cells are not detected, the system moves to the next imaging point.
Artificial intelligence has become commonly accepted in diagnostic imaging and is an increasingly popular tool for a number of applications. Its appeal over traditional mathematical approaches is both its speed and incredible accuracy. However, it is important to be able to validate the results of AI computations, and to utilize these results appropriately for computational analysis.
NIS-Elements software provides feedback during training routines to indicate the confidence of the trained neural network to provide accurate results, as well as several analysis tools and workflows to validate the efficiency of the neural networks, or to allow easy comparison of AI data to ground truth data.
●: included, ⚬: option
| Denoise.ai | Clarify.ai | Enhance.ai | Segment.ai | Convert.ai | |
|---|---|---|---|---|---|
| NIS C, Ar, Ar ML, Ar Passive | Included | Optional | Optional | Optional | Optional |
| NIS.ai Module | No | No | Included | Included | Included |
| Deconvolution Modules | No | Included | No | No | No |
| Batch Deconvolution | No | Included | No | No | No |
| Offline Batch Denoise Package | Included | No | No | No | No |
