28 October 2019

The potential offered by AI in the fields of tumour detection and treatment is considerable. However, machines cannot replace doctors.

“There is a good chance that we will have defeated cancer in 10 to 20 years” – this statement from German Health Minister Jens Spahn at the launch event of the National Decade Against Cancer made the headlines. Several cancer researchers renowned to a greater or lesser extent have also predicted medicine’s triumph over the disease in the past. However, until now these prognoses have not yet turned out to be true. On the contrary, the number of new cases of cancer has almost doubled since the 1970s. At the moment, almost 500 000 Germans per year receive a diagnosis of cancer. As the risk of cancer increases with age and our society is growing older, experts predict an increase to up to 600 000 cases per year by 2030, according to the German Cancer Research Centre (DKFZ). While cancer is the second most common cause of death in Germany today, in other rich countries it has overtaken cardiovascular disease as the most common cause of death among 325 to 70 year olds. This trend is likely to extend to include Germany.

Secret weapon artificial intelligence?

The bad news shown by the statistics also clearly shows the urgency of better cancer research and treatment. For decades, scientists, the pharmaceutical industry and politicians have been working intensively towards both these goals. Thanks to innovative treatment methods and medications, successes have already been achieved. The final breakthrough is expected to be delivered by a special secret weapon in the fight against cancer: artificial intelligence. Prof. Dr. Frederick Klauschen, the medical director of the Institute for Pathology at the Charité Berlin and one of the pioneers in the field, is convinced that “in the next ten years, AI will revolutionise medicine, and therefore also pathology”.

Volumes of data and long analysis times: the challenges of modern medicine

Today, imaging technologies such as X-rays, CT and MRT scans and sonography play an important role in localising tumours during a cancer diagnosis. Another method of diagnosis are biopsies, during which tissue is taken from the area of the body where a tumour is suspected and inspected under a microscope. The images arising from this so-called histopathological diagnosis are evaluated almost entirely manually. “A pathologist evaluates changes to the tissue being examined and decides whether to classify them as pathological or normal variations,” explains Klauschen, who spent five years carrying out research at the National Institute of Health in the USA and was honoured with numerous awards. If small changes are found in cells or tissue architecture, this can indicate a malignant tumour. However, alongside this qualitative assessment, the quantitative evaluation of tissue properties is becoming ever more important, adds the expert, and gives an example: “The quantification of certain protein markers shown by special colours, or the number of immune cells in the tumour”. This is precisely the challenge for modern medicine: “Pathologists can ‘count’ these quantities, but it is either extremely time-consuming or is based on an estimate, with the accompanying inaccuracy”, Klauschen summarises the dilemma. Here, AI methods can be employed. If the computers are trained accordingly, they can recognise tissue structures and changes more quickly and efficiently. This means they can master the most significant challenges facing medical experts today: The immense quantities of data and the resulting analysis times. Instead of running analyses overnight before diagnosis, real-time methods make it possible to carry out the analysis during diagnosis. “Essentially, AI should be advantageous for efficiency and increase diagnosis precision, which also reduces error rates and improves patient safety”, the pathologist summarises the advantages.

^{Prof. Dr. Frederick Klauschen © Charité}

Promsing research results

The first results of the research into artificial intelligence are certainly promising: In 2018, scientists at Eötvös University in Budapest presented an AI-based method, which recognised breast cancer cells from X-ray images of female breasts from mammograms – the usual method of identifying breast cancer – with a probability of approximately 90%. A study by Holger Hänßle from Heidelberg University on early identification of skin cancer also caused a stir. In the study, a computer and an international team of dermatologists simultaneously evaluated identical images of moles and melanoma. The results were remarkable: While the dermatologists identified 86.6% of the malignant melanoma on average, the AI system managed to correctly identify 95%.

At the Charité Berlin, there are also various departments and work groups researching the use of machine learning. One of these is the Department for Radiology: “Quantitative imaging, which also include current AI developments, has been used in radiology for many years,” explains PD Dr. Tobias Penzkofer, head of the work group responsible. “The established techniques are mostly based on the method of analysis of the contrast medium or functional quantitative parameters, such as diffusion imaging. The techniques are used in imaging for diagnosis of liver, prostate or brain tumours, for example.” However, until now it was mostly only possible to observe individual data points; there were no analysis options available for considering large quantities of data. “This will change fundamentally with the new wave of AI-based technologies: We now have the facility to analyse existing image data in a completely new way,” says Penzkofer. Algorithms from the field of artificial intelligence could, for example, find complex systems – not accessible to a human assessor – in depictions of pathological changes to the body. With the exception of individual applications in the field of anatomy identification or automatic labelling of pathological changes, however, these technologies are not yet used in practice. “These newly developed methods are still mostly experimental in character,” notes the medical expert. The same applies to a current research project on non-invasive characterisation of prostate cancer, which the Department of Radiology is working on together with the Department of Urology and the Institute of Pathology. With the aid of artificial intelligence, the aim is to be able to identify malignant tumours in the prostate gland solely from images. The “ambitious goal” of the project, according to Penzhofer, is to save patients the painful, onerous biopsy procedure.

^{Dr. Tobias Penzkofer © Peter Johann Kierzkowski}

Appropriate treatment methods reduce the cost to the healthcare system

Avoiding unnecessary procedures and at the same time reducing costs to the healthcare system is also the goal of another project at the Charité. Together with partners from clinical study groups and the pharmaceutical industry, Prof. Dr. Klauschen is currently developing opportunities for employing AI in the development of medications. “For example, we can search for new biomarkers (Note: Measurable parameters of biological process, which are prognostically or diagnostically significant and can therefore be used as indicators of illnesses, for example), in order to make prognoses about which patients would respond well to a treatment method and which would not,” explains the pathologist. “It can make a statistically significant difference to the evaluation of a medication whether a particular characteristic is demonstrable in, for example, five per cent of tumour cells in a study or in ten per cent of tumour cells. Under a microscope with the naked eye, it is often difficult to quantify this so precisely.” AI, on the other hand, can “read” this type of complex, prognostically or therapeutically relevant property out of the image.

Aignostics: Opening the black box

While this project is still in the early stages, the AI algorithms developed by Klauschen’s team together with Klaus-Robert Müller from TU Berlin in the “Aignostics” project are already being used in daily clinical practice at the Charité. “Typically, an AI system works as a ‘black box’, meaning it is not possible to draw conclusions about why, for example, the AI classifies tissue as a malignant tumour,” describes Prof. Dr. Klauschen the advantage offered by Aignostics, which is supported by the Berlin Institute of Health within the “Digital Health Accelerator” program. “The special thing about our technology is that it illustrates the analysis result with ‘heat maps’, which make it possible for pathologists to verify the AI result at a glance. This difference is essential because the pathologist carries the responsibility for the diagnosis and is not authorised to blindly trust the AI.”

Great potential for molecular pathology

While Aignostics shows the potential offered by AI in histopathological imaging diagnostics, a special method of microscopic study of diseases, machine learning is currently being put through its paces in a clinical study in another field of application within pathology: molecular pathology. “Molecular pathology is playing an ever-greater role in pathological diagnostics, analysing above all genetical changes to tumours,” says Klauschen. “While we have until now analysed individual genes, new methods make it possible to survey tumours’ genetic fingerprints increasingly comprehensively.” At the same time, it is becoming increasingly difficult to interpret this complex data. Here, too, AI methods can help, as researchers from the Charité, the German Cancer Consortium and the TU Berlin recently proved in a study AI methods determined the origin of degenerative tissue using chemical changes to the DNA. As a result, it has been possible to solve a longstanding problem in the diagnosis of patients with head and neck tumours. Some of the 17 000 Germans who develop head and neck tumours each year also develop lung tumours. “In most cases, it is not possible to decisively determine whether this represents the cancer spreading – metastasising – from the head or neck, or whether it is a second tumour – lung cancer,” explains Klauschen, who led the study together with Prof. Dr. David Capper from the Charité’s Institute for Neuropathology, in conversation with analytica-world.com. “For the patient’s treatment, however, this difference is extremely significant. While locally advanced lung cancer can potentially be cured with an operation, patients with a metastasised head or neck tumour have considerably smaller chances of survival and require radiochemotherapy, for example.” Normally, pathologists use analysis of the tumour’s microstructure as well as detection of characteristic proteins in the tissue to distinguish between the two cases. However, in the majority of cases, these analyses do not deliver an unambiguous result. More success has been achieved by analysis of tissue samples with regard to chemical changes to the DNA, so-called methylation. Using methylation data from several hundred head, neck and lung tumours, the researchers trained a deep neural network to differentiate between these types of tumour. “Our neural network is now capable of differentiating lung cancer from metastasis of head and neck tumours with an accuracy of over 99%,” highlights Prof. Dr. Klauschen. “In my opinion, this type of use of AI is just the beginning,” he adds. “We will continue to intensify research here at the intersection of histological and molecular pathology.”

From trials to daily clinical practice

There are several hurdles to be overcome before both the Aignostics algorithms and the above molecular pathological method can be employed for routine diagnostics. The solutions require CE certification, for example – new ground in the field of machine learning. The core requirement for implementation, in Klauschen’s opinion, remains the willingness of institutions to make the significant investments connected to digitalisation. In addition, data protection rules are required, as well as a uniform approach to testing the quality of AI algorithms from a regulatory point of view. “I am part of an initiative from the standardisation organisation ITU of the United Nations in cooperation with the World Health Organisation to develop standards for validating AI methods for medical applications, and we need to clarify to what extent we are permitted to use patient data to train AI methods, including in cooperation with companies.” These are difficult topics, on which the various institutes of the Charité which are involved in the fields of imaging and AI regularly exchange information. Just as they do on another important prerequisite for successful implementation of AI: “Last but not least, the solution must be able to be effectively embedded into the work processes and system landscape of each institute,” says Prof. Dr. Klauschen, who also leads the research field of system pathology at the Charité. “Of course, this requires a series of interfaces, which we are still in the process of developing.”

The experts from all disciplines at the university clinic agree on one thing: As great as the potential offered by AI is, it will not take over the doctors’ work. “Current results indicate that the greatest potential is currently in “supporting AI” for radiologists,” suspects Dr. Penzkofer. “Here, repetitive, error-prone tasks will be delegated to AI, improving the quality of diagnoses. These tasks also include quality control and planning of examinations, as this allows the strengths of the untiring algorithms to be used”. In addition, AI could take over uninteresting, labour-intensive tasks such as searching for metastasis in lymph nodes, confirms Dr. Penzkofer’s pathologist colleague. “Pathologists will not become superfluous, rather AI will be used as a diagnostic aid, to handle increasing numbers of cases and complexity. Like every new diagnostic technique, we need experienced experts to interpret the test results in a clinical context,” believes Prof. Dr. Klauschen. “Artificial intelligence will allow us to concentrate on more important things, rather than tasks which the computer is better at.”