Artificial intelligence is currently one of the most innovative issues, but also one of the most controversial research areas. It already has a firm footing in many areas of our everyday life. Whether it is the automotive industry or online marketing, artificial intelligence is already being used, and often we are not even aware of it. Artificial intelligence has long been an integral part of many processes in research and diagnostics in medicine and the life sciences – and it will be even more widely used in the future.
Increasing automation in laboratories and clinics, and the development of high-throughput techniques have led to a flood of data in the life sciences in recent years. The ability to analyse many parameters in the shortest possible time has made possible many innovative developments that benefit people and the environment. However, a large portion of the data produced has often been left unused because evaluating all this data has presented a virtually insurmountable challenge. IT specialists have therefore begun to develop algorithms with human-like decision structures so that the data can be used as comprehensively as possible. This extremely complex area of computer science is commonly referred to as "artificial intelligence (AI)".
Economist Sergey Biniaminov, who is head of the Karlsruhe-based company HS Analysis, is a specialist in this field. The company is specialised in the management of large data volumes and software infrastructures in the life sciences. “For me, AI means that an algorithm is capable of making decisions of a certain complexity independently,” he explains. “A sub-area of the life sciences plays a particular role – machine learning, which helps process information such that the algorithm is able to carry out actions autonomously according to a given procedure. However, as the application of machine learning is so diverse, we generally find it very difficult to come up with precise definitions in our field.”
Of the many machine learning methods, a method known as deep learning is often used in the life sciences. It is an algorithm that enables autonomous, machine-based generation of knowledge and experience. “Human knowledge is based on a wealth of information. This is how we build experience, and it is precisely this information that serves as data input for the machines that collect information, comparing new data with old and subsequently making decisions on their own and learning from this,” says Biniaminov, explaining deep learning. “We are experiencing a real boom at the moment, and deep learning is being used sustainably in various disciplines. So much is happening these days because we now have the technical capabilities.”
At HS Analysis, many projects involve the deep learning method. One of the company’s business areas is focused on evaluating extensive data from microscopic images. "Deep learning creates new opportunities to recognise and correlate individual relationships and objects in microscopic data, enabling new insights," says Biniaminov.
The roots of AI, like automation, are actually to be found in the automotive industry, which is carrying out intensive research aimed at developing such data analyses for autonomous driving. Just as Henry Ford brought automation to the lab decades after automation had first been applied in the automotive industry, the automotive industry initially also used a promising type of deep learning called convolutional neural networks (CNN) for image processing. The methods were subsequently used in other business sectors as well. Nowadays, AI is applied in an almost unimaginable number of ways in the fields of medicine and life sciences.
AI methods have long become standard in medical diagnostics. For example, they can be used for examining membranous glomerulonephritis, one of the most common renal diseases in adults, in order to identify, quantify and compare with disease data of individual kidney objects in a highly reliable way. “This can also be transferred to other organs and their segments. More precise diagnoses enable the treatment of a broad range of diseases,” explains Biniaminov, who is currently involved in establishing a Germany-wide diagnostics network.
That said, AI has its limitations and weaknesses, just like any other technology. AI has come under scrutiny from sceptics, who fear that it will be impossible to handle all the data in the international race to becoming number one in this technology. "We make sure that we always communicate the limitations of the technology. This is the only way of solving potential problems,” says the CEO. “Software always has a probability of error. This is why we cannot just apply the general models to any task whatsoever; there is no one-size-fits-all solution. You have to rely on individual solutions to perform certain AI-powered tasks.”
Biniaminov further explains that highly accurate results can be obtained, but in the majority of cases it is difficult to figure out exactly why a model produces a certain result. He comments: “Our mission is to explain why the models have come to a certain conclusion, but as things stand at the moment, this is still a very difficult task.” He says that this is partly because of CNNs, convolutional neural networks, on which AI is based. There are millions of decision levels and possibilities. Therefore, as things currently stand, humans are unable to interpret the results.
Biniaminov is convinced that despite the limitations and weaknesses of the new technology, the benefits will still greatly outweigh the disadvantages. He comments: "Of course, humans are excellent at analysing things if they look at just a single image. But if thousands of images or features have to be evaluated, then a robot or an algorithm can perfectly well support humans in making decisions. That's why humans need machines as assistants."
Biniaminov also believes that humans and AI assistants need to be able to communicate with each other in order to compensate for the disadvantages of artificial learning. He is currently working on an optimal human-machine collaboration. "The vision is to combine several different neural networks. These could then make decisions independently like two people. One network would deliver the decisions of another network in a logic at a dimension understandable for humans," he explains. "All we have to do is visualise it and bring it into the diagnostic context. In order to achieve this, it is important to involve specialists and check reproducibility and objectivity in concrete projects.”