Artificial intelligence always triggers a lot of curiosity when mentioned around a dinner table. AI brings in mind to a lot of people the latest fantastic movie released in the cinema with cool robots and futuristic computer systems. In reality and up until today, AI is really the science of giving to a material device the ability to show cognitive behaviour similarly to humans in order to learn and solve complex problems. AI is very present in our everyday lives and its development and achievements made in various fields is increasing efficiency and accuracy of many complicated tasks.
In this article we will discuss in particular how AI is applied to automation and zebrafish egg sorting.
Artificial intelligence
Artificial intelligence (AI) is the capacity of a machine to independently perceive its environment and act accordingly to achieve a specific goal. In other words, it is a computerised system capable of interpreting, understanding and processing external data and generating a sensible output. Intelligent devices are highly efficient in solving complicated tasks and problems by integrating large quantities of data which would be too strenuous to handle manually.
The main goal of artificial intelligence is to develop a technology allowing computerised devices to work in an intelligent and cognitive way. However, this is an extremely challenging task. To build an intelligent device, the technology must be developed based on several essential features. (1)
Dealing with uncertain and incomplete information to solve problems and make logical reasonings.
Describing objects, relations, concepts and properties for them to be interpreted by software agents.
Setting goals and achieving them.
Improving problem solving through experience.
Reading and understanding human language.
Use inputted information from sensors such as sonars, radars, microphones, and so on to deduce aspects from the surrounding environment.
Moving efficiently in a small, static and visible environment as well as recognising locations.
Recognising, interpreting, processing and reproducing human emotions, actions and interactions.
Reproducing human cognition.
In reality, the general term of artificial intelligence is very broad and needs to be specified according to the context in which it is applied. Thus, many subfields have been described such as robotics and machine learning and are distinguished based on the different tools and intelligent features they use.
As a matter of fact, AI is currently being developed as an innovative approach for drug design and health care systems. For example, the pharmaceutical industry is showing wide interest in deep learning architectures to predict properties and activities of chemical compounds (2). AI has also made its way into the medical field where complex medical cases require high levels of knowledge and data analysis to be solved. Medical artificial intelligence has the intent to help clinicians formulate a diagnosis, take a therapeutically decision and predict a probable outcome (3).
Machine Learning for zebrafish classification
Our interest lies in machine learning as it is the basis to our algorithms for egg classification and sorting. Machine learning enables computerised systems to build and refine models for a prescribed task or situation. It is a well-established technology for learning and predicting specific outcomes. Learning can be supervised or unsupervised according to the human contribution to the task. While supervised learning requires human based description of the input data (ground-truth labels) to achieve classification or numerical regression, unsupervised learning will recognise patterns that capture the characteristics of the data without a priori knowledge of the expected outcome (1).
Machine learning has already proven to be very popular for automated screenings on zebrafish. As discussed in our previous article, automating as many steps as possible when working with zebrafish is essential to gain time and achieve reliable results. Therefore, setting up an intelligent system to collect, process and screen data to extract features and classify individual zebrafish eggs, embryos or larvae is at the heart of automation.
For instance, an algorithm for sex classification of zebrafish based on their sexual dimorphism has been thought and successfully developed taking advantage of machine learning techniques. Deep Convolution Neural Networks (DCNN) and Support Vector Machine (SVM) are combined to determine the zebrafish sex based respectively on the body colour and pattern and the caudal fin colour. Those learning methods are trained with image inputs to classify zebrafish according to their gender and are uniquely based on image analysis (4).
Regarding zebrafish embryos, various software programs have been developed to classify them through pattern recognition of features detected on images. Researchers at KIT (Karlsruhe Institute of Technology) have developed a classification software able to differentiate wild-type embryos from mutant individuals with an accuracy of 79-99% depending on the test and system. Their system enables to extract phenotypic features and convert them to a numeric representation for further classification using SVM (5).
Although several machine learning approaches are available and can achieve reliable results, Deep Neural Networks (DNN) is the most popular system exploited so far. The following toxicology study reports a fully automated machine learning-based approach to separate healthy from dead (coagulated) embryos with an accuracy rate reaching up to 99.47% (6). This study is highly related and relevant to our technology.
Figure 1: Zebrafish egg images. (a) and (b) are live eggs, whereas (c) and (d) are dead and coagulated eggs.
As mentioned above, the intelligent system operated by the EggSorter is as machine learning-based approach, which we will further discuss in the following section.
Classification algorithms in the EggSorter
The EggSorter software has been developed to classify eggs according to fertilisation and fluorescent markers. Using machine learning algorithms, the EggSorter is capable of recognizing different morphological features on the eggs with high accuracy. As an example, it can tell if an egg is fertilized or not from 1 hpf with a false discovery rate less than 5%, if the classification is done from 1 hpf to 2 hpf, and less than 1% if done from 2 hpf to 5 hpf.
Furthermore, some other machine learning classification algorithms implemented in the EggSorter are able to tell whether the eggs are fluorescent or not. This is of great importance when creating new lines or when selecting the eggs to prepare them for the different experiments.
Although we are now capable of using our device for egg sorting according to both of the previously mentioned traits, we will implement in the future additional sorting and classification features, which could be personalised to a specific experiment.
Share if you liked it!
References
(1) Russell, S. J., Norvig, P., & Davis, E. (2010). Artificial intelligence: a modern approach. 3rd ed. Upper Saddle River, NJ: Prentice Hall.
(2) Hessler, G., & Baringhaus, K. H. (2018). Artificial Intelligence in Drug Design. Molecules (Basel, Switzerland), 23(10), 2520. https://doi.org/10.3390/molecules23102520
(3) Ramesh, A. N., Kambhampati, C., Monson, J. R., & Drew, P. J. (2004). Artificial intelligence in medicine. Annals of the Royal College of Surgeons of England, 86(5), 334–338. https://doi.org/10.1308/147870804290
(4) Hosseini, S., Simianer, H., Tetens, J., Brenig, B., Herzog, S., & Sharifi, A. R. (2019). Efficient phenotypic sex classification of zebrafish using machine learning methods. Ecology and evolution, 9(23), 13332–13343. https://doi.org/10.1002/ece3.5788
(5) Schutera, M., Dickmeis, T., Mione, M., Peravali, R., Marcato, D., Reischl, M., Mikut, R., & Pylatiuk, C. (2016). Automated phenotype pattern recognition of zebrafish for high-throughput screening. Bioengineered, 7(4), 261–265. https://doi.org/10.1080/21655979.2016.1197710
(6) Tharwat, A., Gaber, T., Fouad, M. M., Snasel, V., Hassanien, A. E. (2015). Procedia Computer Science, 65. Towards an Automated Zebrafish-based Toxicity Test Model Using Machine Learning, International Conference on Communication, management, and Information technology (ICCMIT’2015).