You finish tinkering with several parameters and click analyze to test them on your image. You hold your breath as the traditional image analysis software runs analysis pipelines and extracts the features of interest. You look over the results of this test run with apprehension, but your blood pressure starts to lower as you see that they finally look accurate. You move to the next image showcasing cells that were treated with one of the drugs from a library and optimistically click analyze. Your heart sinks as you realize the results look nothing like the first image and you now must reset the parameters to fit both images.

The challenge of optimizing analysis parameters is nothing new to the image analysis process, but one that many researchers would like to avoid. Given the importance of image analysis in many assays needed for drug discovery, it is critical to use better methods that can save time, money and frustration. Fortunately, the headache of manually optimizing image analysis has become increasingly avoidable with the development of deep learning capabilities.

Getting it right

Deep learning, which is one of the methods in artificial intelligence (AI), is proficient at setting the parameters for image analysis so researchers do not have to endlessly adjust them. While the deep learning algorithm must be trained, the process is simple. Researchers need only trace digital boundaries around objects of interest – alongside any background noise they wish to ignore – in a set of example images, thereby teaching the algorithm to find those same features in similar images and bypass the rest.

Because variability is a common concern present in all image analysis methods, introducing the algorithm to a diverse set of training images that range in complexity is of critical importance. The more variety it encounters during the learning phase, the more accurately it can adjust to variability between images that are bound to appear.

This action creates a deep learning model that has captured your teachings and is empowered to effectively implement that knowledge when analyzing hundreds, thousands or millions of other images going forward.

Dashboard seen in IN Carta® Image Analysis Software. Credit: Molecular Devices.

Addressing time, quality and cost

Deep learning makes image analysis more hands-off for researchers offering several notable advantages. The first is time savings. Instead of investing hours into the tedious process of optimizing image analysis parameters, researchers train the algorithm that generates a deep learning model used to analyze images – and the computer takes over from there. This increased freedom ultimately gives researchers more walkaway time that can be spent interpreting results and tackling other critical tasks.

Deep learning can also improve the quality of the image analysis being done. When performed manually, some researchers eventually reach burnout and do not spend the time needed to optimize their analysis parameters, instead resorting to use of suboptimal parameters that somewhat work for their image set. Using deep learning would greatly improve the work being done in these cases.

Additionally, turning image analysis over to deep learning can save researchers money. By streamlining the process, researchers can scale up their image analysis efforts while creating more reliable results with less work. More robust results derived from high-content imaging can lead researchers to quickly make conclusions and devise a plan for the next research phase, with increased confidence in their image analysis.

Beyond time savings, improved quality and reduced costs, deep learning also allows for improved analysis of some challenging assays, like label-free, which were previously difficult to develop a robust algorithm for.

Demystifying AI

While deep learning has brought significant improvements to image analysis, relieving a heavy burden for researchers, it is still seen as mysterious in some circles – and understandably so. It is difficult to truly know what the algorithm learned, how it learned and why it analyzed images the way that it did. There is also an assumption that researchers need a computer science background to incorporate deep learning into their routine experiments.

However, with innovative, user-friendly software programs that exist today, virtually anyone can learn to use them without needing to understand the complex world of neural networks, that more advanced models require. Researchers can achieve reliable, transformational results compared to the traditional analysis approach.

Customized deep-learning models segment specific regions (whole body, head, eyes, brain) of a zebrafish embryo in label-free images. Models have been trained in IN Carta SINAP. Credit: Image courtesy of Guo Lab, UCSF.

Looking to the future

Image analysis is a crucial part of the drug discovery process. But today, it is one which requires a lot of resources and energy from researchers to optimize each individual assay. Integrating deep learning algorithms into the analysis can streamline the process, provide more robust results and save researchers from tediously adjusting analysis parameters. When provided with diverse training images, deep learning algorithms can take the frustration out of image analysis and provide quality results even for complex in vitro models like organoids. Integrating deep learning into image analysis will accelerate the drug discovery process and enable researchers to get more done – with fewer headaches.

While deep learning has elevated image analysis from the time-consuming manual process to the optimized route requiring less guidance, the experience will be even simpler in the future. Instead of having to manually circle objects of interest for training, future models will allow users to simply click on a few representative biological structures. The software could turn into a robo-assistant and provide several segmentation suggestions, from which the user can pick the most accurate one. This simplified user experience, combined with a greater ability to detect biological objects, will allow for maximum output and time savings.

Recent posts