Artificial intelligence helps diagnose leukemia
Artificial intelligence helps diagnose leukemia
Researchers at the National Center for Tumor Diseases Dresden (NCT/UCC), the University Hospital Carl Gustav Carus Dresden and the TU Dresden have for the first time developed an artificial intelligence (AI)-based computer system that can provide highly accurate support in the initial diagnosis of acute myeloid leukemia (AML). In addition, it can predict in the vast majority of cases a mutation that is important for the disease. The system is based on artificial neural networks, which the scientists trained using digital image data from bone marrow smears. In the future, the AI-based analysis software should support physicians in making diagnoses in everyday medical practice. The results of the study were published in the journal Leukemia (doi10.1038/s41375-021-01408).
The National Center for Tumor Diseases Dresden (NCT/UCC) is a joint institution of the German Cancer Research Center (DKFZ), the University Hospital Carl Gustav Carus Dresden, Carl Gustav Carus Faculty of Medicine at TU Dresden and the Helmholtz-Zentrum Dresden-Rossendorf (HZDR).
Acute myeloid leukemia (AML) is the most common type of rapidly progressing blood cancer in Germany. If left untreated, it inevitably leads to death, and even with the most modern therapies, a cure is only possible for a minority of patients. Precise diagnostics and a timely start of the best possible therapy are therefore extremely important. If AML is suspected, analysis of a bone marrow smear provides a quick overview. "However, the evaluation of microscopic images is highly complex and has so far been highly dependent on the experience of the respective physician. Our work shows that artificial intelligence is capable of supporting experts in their difficult task by providing a purely data-based, objective finding," explains Prof. Martin Bornhäuser, member of the Directorate of the National Center for Tumor Diseases Dresden (NCT/UCC) and Director of Medical Clinic I at the University Hospital Carl Gustav Carus Dresden.
For the first time, scientists at the NCT/UCC, Dresden University Hospital, TU Dresden and the Mildred Scheel Early Career Center, with the support of the Else Kröner Fresenius Center (EKFZ) for Digital Health, have developed an AI-based computer system that can distinguish between the bone marrow smear of an AML patient and the sample of a healthy person with over 95 percent discriminatory power.
For this purpose, the researchers used artificial neural networks, a subfield of artificial intelligence that mimics humans' ability to learn from examples. Large amounts of data are needed to train the neural networks. In this case, the system first learned to recognize and distinguish cells from each other using digitized image data of bone marrow smears from 1,251 AML patients and 236 healthy bone marrow donors. Experts manually outlined more than 90,000 individual cells as the basis for the machine learning process. The computer was then trained to distinguish between different cell types and characteristics. "Artificial neural networks are able to analyze and quantify a large number of features very quickly. On this basis, our system can very precisely distinguish smears from AML patients and healthy individuals," explains Dr. Karsten Wendt from the Institute of Software and Multimedia Technology at TU Dresden.
Artificial intelligence predicts mutation based on image features
The software is also able to predict a specific genetic change – a mutation of the gene nucleosphosmin (NPM1) – with an accuracy of over 85 percent based on external cell characteristics. The mutation, present in about a quarter of adult AML patients, is associated with a comparatively good prognosis and is important for selecting the appropriate therapy. "We were able to show that the AI-based analysis software uses previously unknown morphological features to detect the mutation, which also provide clues to as yet unexplored relationships in cell biology," says co-project leader Dr. Jan-Niklas Eckardt. Despite the excellent results, the system is not intended to replace the assessment by specialists in the future, but to provide them with useful support for a rapid initial diagnosis. Additional, more time-consuming procedures such as genetic analysis will also remain indispensable for the final diagnosis of AML in the future.
To use the newly developed software solution, laboratory technicians or physicians only have to select relevant image areas from the bone marrow smears and feed the digital image data into the system. "All further analysis steps, such as categorizing and counting cells, which have been done by hand over the past 50 years, are carried out fully automatically," explains Dr. Jan Moritz Middeke, who heads the working group together with Dr. Karsten Wendt. The technological approach serves as a groundwork for many other image-based examination methods and is to be extended to more complicated applications.
Publication:
Eckardt, JN., Middeke, J.M., Riechert, S. et al. Deep learning detects acute myeloid leukemia and predicts NPM1 mutation status from bone marrow smears. Leukemia (2021). https://doi.org/10.1038/s41375-021-01408-w
Press Contact:
Dr. Anna Kraft
Nationales Centrum für Tumorerkrankungen Dresden (NCT/UCC)
Presse- und Öffentlichkeitsarbeit
Tel.: +49 (0)351 458-5548
E-Mail: anna.kraft@nct-dresden.de
www.nct-dresden.de