AI against cancer

The prick of a needle, a few drops of blood – and a computer. In the future, this may be all that is needed to diagnose cancer. A new approach from the research group led by G.V. Shivashankar – head of the Laboratory for Multiscale Biology in the PSI Center for Life Sciences and professor of mechanogenomics at ETH Zurich – aims to make this possible. The method combines modern imaging, cell biology, and adaptive algorithms, a combination that turns the blood sample into a valuable source of information about a possible cancer.

How tumours leave traces in the blood

Tumour cells release certain messenger substances into the bloodstream – a complex mixture known as the secretome. It differs depending on the type of tumour, and it affects distant cells in the body. The researchers at PSI determined that this secretome influences white blood cells in particular – especially their nuclear structure. This is where chromatin, genetic material coiled tightly like a ball of wool, is located. “White blood cells react very sensitively to such tumour signals,” Shivashankar explains. “Because they are very easy to obtain from blood samples, an analysis proves to be extremely straightforward.”

But exactly what changes are taking place in the chromatin can’t be seen with the naked eye – or even directly with a microscope. The differences are subtle and complex, and they cannot be attributed to a single characteristic. Also, the pattern of change is heavily dependent on the type of tumour.

To decipher these invisible signatures, the researchers used high-resolution fluorescence microscopy. They generated several hundred images per blood sample and analysed around 200 properties – such as texture, contrast, and spatial distribution of the chromatin. But even with these images, the question remained: Which combination of characteristics actually indicates cancer – and if so, what type? To draw a reliable conclusion from this complexity, the researchers are relying on artificial intelligence – specifically, machine learning.

How an algorithm learns from images of cells

To be able to draw reliable conclusions from the multitude of microscopic features, the research team used a learning algorithm. First they analysed blood samples from healthy individuals and patients with different types of tumours. The resulting cell images formed the basis for training the algorithm: Each image shows a chromatin structure – along with information indicating whether the corresponding individual is healthy or afflicted with cancer.

“You can think of it as being like a card game,” Shivashankar explains. “One side of each card shows the image of the cell – and on the other side is the diagnosis.” During the training phase, the algorithm is permitted to see both sides. It compares hundreds of such “cards” and begins to recognise characteristic patterns – including those that are invisible to the human eye or are ambiguous.

Learning, in this case, means: The algorithm assigns different weights to the microscopic abnormalities – and as the amount of data increases, it increasingly “understands” which combination indicates cancer. Then, during the testing phase, it is dealt with new cards whose “backs” are unfamiliar. It has to decide for itself whether the image indicates cancer – or not.

The result: With an accuracy of 85 percent, the algorithm was able to correctly distinguish healthy individuals from those suffering from cancer – regardless of the type of tumour. “This is the first time worldwide that we have been able to use images of blood cell chromatin to identify tumour markers,” Shivashankar says. This is a promising step towards data-based diagnostics that could, in the future, provide targeted support to medical professionals in early detection – efficiently and cost-effectively. According to the researchers, initial clinical applications should be possible within three to five years.