Computational Biology AI Transforms Blood Disease Diagnosis

The conventional diagnostic method is as monotonous as it is tedious: When blood diseases are suspected, a patient's blood sample arrives at the laboratory, where the cell structure becomes visible under the microscope - and specially trained experts look for distinctive patterns in it, such as those exhibited by leukemia. "In large laboratories, this takes place thousands of times a day," says Dr. Carsten Marr of Helmholtz Munich. He heads a research group that aims to automate this diagnosis with the help of artificial intelligence - and may even be able to deliver more precise results in the process.

There are two basic methods for blood tests: The first is genetic testing, which looks for known mutations in the laboratory. The second method is morphological examination. This involves a so-called blood smear - that is, a drop of peripheral blood is placed on a slide, smear out and stained. Specially trained experts - known as cytologists – use the microscope to see if there are any abnormalities in cells. In certain forms of leukemia, for example, special rod structures appear in the white blood cells; this, along with other features, is an unmistakable sign of the disease. "Some morphological characteristics of leukemia are well-defined and clearly recognizable," says Carsten Marr, "so the field is tailor-made for the use of artificial intelligence."

To train the computer, Marr works with the Munich Leukemia Laboratory (MLL), where thousands of samples come together every day, some of which are digitized. Marr and his team feed the computer images of blood from patients with the disease, as well as images of healthy blood cells. The artificial neural networks analyze these differences and memorize typical patterns, so to speak. This is the classic learning process used to train artificial intelligence. In the meantime, the computer can independently identify diseased cells on this basis.

Reliability is one of the key criteria that must be met if the technology is to be used in practice. A 75 percent hit rate is not enough. However, for other forms of leukemia - such as acute myeloid leukemia (AML) - the hit rate is at least as high as for human cytologists. This is precisely where the next challenge awaits Carsten Marr's specialists: they want to know exactly how the artificial intelligence arrives at its diagnosis. Experts often refer to a "black box" in this context: You feed the computer with photos that it is supposed to analyze independently - and afterwards, you often don't know on the basis of which characteristics the distinctions are made. "It could be, for example," explains Carsten Marr, "that you show the computer the patients with a certain disease based on five-year-old images, but the healthy comparison group is based exclusively on more recent images. The images could differ in color intensity, for example, because of their different ages, even though the human eye doesn't perceive these differences at all." The computer would therefore define the difference between patients based on the color intensity of the images, not on disease characteristics - and the classification would be flawed.

To prevent exactly such mistakes, researchers in artificial intelligence want to break through this black-box character. In plain language, they want to know exactly how the computer arrives at its results. "In our case, it turned out that the artificial intelligence paid attention to the same cells as a human cytologist. We are currently investigating whether this also applies to morphological features, such as the size of the cell, its shape or texture," says Carsten Marr. Another challenge is the generalizability of the models: So far, all of the samples with which the artificial intelligence has been trained have come from the same laboratory, which always works with the exact same procedure and the same technique. If data from other labs differ in color, sharpness or size, a human cytologist can handle it without problems, but the technology fails. "We have to train the artificial intelligence with heterogeneous data sets from different labs so that it can manage this abstraction, or improve the algorithms" says Carsten Marr - a difficult undertaking for which he recently held a hackathon with students: Various teams pounced on the problem and spent a week working intensively to find the best solutions.

For the researchers, it is clear that the effort is worthwhile: the methods with which artificial intelligence searches for diseases can also be extended to other applications. It is conceivable that blood cells could be analyzed in the search for other forms of leukemia or even for completely different diseases. What's more, Carsten Marr is already working on having images of bone marrow examined by the computer as well. "This is much more complex because the cells in the bone marrow often clump together. They lie in closer clusters together and sometimes on top of each other, and there are many more blood cell types," says Marr, outlining the challenge.

The chances that he and his team will be successful in this next project are good: In the field of optical analyses of single cells using artificial intelligence, the experts at Helmholtz Munich are among the world's leaders.

Latest update: March 2023.

Watch Marr's Videos