Using information from tumour cells: AI in medical diagnostics

A group led by G.V. Shivashankar, head of mechano-genomics at PSI and professor at ETH Zurich, was able to prove that changes in the organisation of the cell nucleus of some blood cells can provide a reliable indication of a tumour in the body. With their technique - using artificial intelligence - the scientists were able to distinguish between healthy and diseased people with an accuracy of around 85 per cent. They were also able to correctly determine the type of tumour disease - melanoma, glioma or head and neck tumour. "This is the first time in the world that anyone has managed to do this," says Shivashankar. The researchers have published their results in the journal npj Precision Oncology.

Indicators for tumour cells

Detecting cancer in the body or monitoring the course of treatment is usually very time-consuming and often only takes place at a late stage when the signs are becoming more obvious. Basic researchers are therefore looking for methods that are both easy to use in everyday clinical practice and reliable and sensitive. Shivashankar's research group focussed on lymphocytes and monocytes, which are known in specialist circles as mononuclear cells of the peripheral blood. These are easy to obtain from a simple blood sample and have a round nucleus that is clearly visible under the microscope. According to the researchers' assumption, the genetic material located there reacts to substances in the bloodstream that the tumour releases, the so-called secretome. This activates the chromatin in the nuclei of the blood cells, thus changing the organisation of the genetic material within them. This in turn can serve as an indicator or biomarker. "Our hypothesis was that the blood cells are tumour detectors - and that took us a long way," explains Shivashankar.

Artificial intelligence as a diagnostic aid

The researchers used fluorescence microscopy to analyse the chromatin of blood cells, which is what the genetic material DNA is packaged into a kind of knot. In doing so, they recorded the texture, the packing density or the contrast of the chromatin in the lymphocytes or monocytes, totalling around two hundred characteristics. They fed the microscopy images of healthy and diseased test participants into an artificial intelligence (AI) system. They used the conditions of supervised learning, which serve to teach the software known differences. In the subsequent deep learning approach, the algorithm then identified differences between "healthy" and "diseased" cells that are not recognisable to the human observer.

The research group pursued three different approaches. In a first series of tests, they investigated whether the method can differentiate between healthy controls and patients. To do this, they compared the blood cells of ten patients with those of ten healthy individuals. The AI was able to distinguish between healthy and cancer patients with an accuracy of 85 per cent. "Even analysing just one random cell was still very accurate," says Shivashankar. In a second approach, the aim was to determine whether the AI could even distinguish between different types of tumour. To do this, the researchers fed the algorithm with the chromatin data of the blood cells of ten patients each with a glioma (tumour of the stalk tissue of the nerve cells), a meningioma (tumour of the meninges) and an ear, nose and throat tumour. This test also proved to be successful. The assignments were more than 85 per cent accurate. Finally, a third question concerned patients who underwent or had undergone radiotherapy at the PSI Centre for Proton Therapy ZPT.

Damien Weber, Director and Chief Physician of the ZPT, sees great potential in the diagnostic approach. He asked 150 of his patients to consent to their blood samples being analysed for the study: "We hope that the new method will improve both the diagnosis and the monitoring of the success of the therapy."

In order to record the success of the intervention, blood samples were taken before, during and after radiotherapy. Here, too, the software worked successfully and correctly assigned the samples with a very high degree of accuracy. The treatment was expected to reduce the concentration and composition of tumour signals in the blood - and so it did, and the appearance of the genetic material of the blood cells normalised. "It was amazing to observe how the structure of the chromatin returned to a more healthy pattern over the course of the treatment," said Shivashankar with satisfaction.

Many applications conceivable

According to the biologist and his colleagues, the new method based on blood cell chromatin, which can currently only be used for research purposes, can be applied not only to the tumours being investigated, but also to numerous types of cancer. And it could be limited not only to monitoring the progress of proton therapy, but also to many other forms of therapy, such as radiotherapy in general, chemotherapy and surgery. Further research will now have to prove whether this is the case. Shivashankar's group, in collaboration with the Centre for Radiopharmaceutical Sciences (CRS) at PSI, has already tested whether the chromatin biomarkers can be used to detect radioresistant and chemoresistant cells, as published in the journal Scientific Reports. There is still a lot of work to be done before the procedure can be authorised by the authorities in clinical practice - especially studies with a larger number of participants in order to clarify how high the number of false positive alarms and false negative statements are under clinical conditions. There is no doubt in Shivashankar's mind that the path to clinical application is clear and that patients will benefit from the method. "The method is ready," he says.

Article from "LABO" from 20/02/2024